Theoretical Aspects of Gitksan Phonology

by

Jason Camy Brown

B.A., California State University, Fresno, 2000 M.A., California State University, Fresno, 2002

A THESIS SUBMITTED IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF

DOCTOR OF PHILOSOPHY

in

The Faculty of Graduate Studies

(Linguistics)

THE UNIVERSITY OF BRITISH COLUMBIA (Vancouver)

December 2008

© Jason Camy Brown, 2008 Abstract

This thesis deals with the phonology of Gitksan, a Tsimshianic language spoken in northern British Columbia, Canada. The claim of this thesis is that Gitksan exhibits several gradient phonological restrictions on consonantal cooccurrence that hold over the lexicon. There is a gradient restriction on homorganic consonants, and within homorganic pairs, there is a gradient restriction on major class and manner features. It is claimed that these restrictions are due to a generalized OCP effect in the grammar, and that this effect can be relativized to subsidiary features, such as place, manner, etc. It is argued that these types of effects are best analyzed with the system of weighted constraints employed in Harmonic Grammar (Legendre et al. 1990, Smolensky & Legendre 2006). It is also claimed that Gitksan exhibits a gradient assimilatory effect among specific consonants. This type of effect is rare, and is unexpected given the general conditions of dissimilation. One such effect is the frequency of both pulmonic pairs of consonants and ejective pairs of consonants, which occur at rates higher than expected by chance. Another is the occurrence of uvular-uvular and velar-velar pairs of consonants, which also occur at rates higher than chance. This pattern is somewhat surprising, as there is a gradient prohibition on cooccurring pairs of dorsal consonants. These assimilatory patterns are analyzed using the Agreement by Correspondence approach (Hansson 2001, Rose & Walker 2004), which mandates that output correspondents agree for some phonological feature. The general discussion of assimilation and dissimilation is continued in morphological contexts, such as reduplication. It is claimed there are differences in the gradient and categorical patterns of assimilation and dissimilation in Coast Tsimshian and Gitksan reduplicative contexts. A summary of the attested reduplicative patterns in the languages, as well as results from a nonce-probe task, supports this claim. This difference between Coast Tsimshian and Gitksan is indicative of a larger difference in the reduplicative patterns of the languages of the Tsimshianic family: each member of the family exhibits slightly different patterns of deglottalization. A typological study of these patterns suggests that glottalized sonorants and obstruents are fundamentally different segment types.

11 Table of Contents

Abstract i Abstract ii Table of Contents iii Acknowledgments v Dedication viii Chapter 1: Introduction 1

1.1. Introduction 1 1.2. Overview of Gitksan 4 1.2.1. Language Area 4 1.2.2. The Tsimshianic Language Family 5 1.2.3. Interior Tsimshianic 7 1.2.4. History and Documentation 8 1.2.5 The Gitksan Database 9 1.2.6. Orthography 10 1.3. Theoretical Overview 11 1.3.1. Documenation and Theory 11 1.3.2. Constraint-Based Theories and Gradient Phenomena 12 1.4. Structure of the Thesis 13 Chapter 2: Patterns Across the Lexicon 15 2.1. Introduction 15 2.2. Consonantal Patterns Over the Gitksan Lexicon 16 2.2.1. Overview and Method 16 2.2.2. Results 29 2.2.2.1. Gradient OCP Effects 29 2.2.2.2. Gradient Assimilatory Effects 39 2.3. Summary. and Discussion 48 Chapter 3: Constraints on Gradient Lexical Phonotactics 50 3.1. Introduction 50 3.2. Background on Theoretical Approaches 50 3.2.1. Problems for Optimality Theory 52 3.2.2. Lexical Indexation 54 3.2.3. Weighted Constraints 57 3.3. Harmonic Grammar 58 3.3.1. Overview 58 3.3.2. Constraints 63 3.3.3. Learning the Grammar 66 3.4. Distance in OCP Effects 76 3.5. Conclusion 78 Chapter 4: Laryngeal Features 79 4.1. Introduction 79 4.2. Glottalization Agreement 82 4.2.1. Analysis 87 4.3. Paradigmatic Restrictions on Laryngeal Features 102

111 4.3.1 Obstruent 103 4.3.2. Blocking allophonicvoicing. voicing 109 4.3.2.1. Ejectives 109 4.3.2.2. Aspirated Stops 110 4.3.2.3. Fricatives 112 4.3.2.4. Summary of Blocking 113 4.4. Conclusion 114 Chapter 5: Dorsal and Guttural Patterns 115 5.1. Introduction 115 5.2. Lexical Patterns 116 5.3. Arguments for Dorsal vs. Pharyngeal Place 119 5.3.1. Guttural Lowering 124 5.3.1. Constraints on Guttural Lowering 134 5.4. Analysis of Lexical Patterns 138 5.5. Conclusion 145 Chapter 6: Reduplicative Patterns 146 6.1. Introduction 146 6.2. Reduplication Basics 146 6.3. Gradient Reduplicative Patterns 149 6.3.1. Frequencies of Plural Morphology in Tsimshianic 151 6.3.2. Reduplicative Allomorphy in Coast Tsimshian 154 6.3.3. Reduplicative Allomorphy in Gitksan 156 6.3.4 Testing Reduplicative Frequencies 162 6.3.4.1 Stimuli 162 6.3.4.2. Subjects 163 6.3.4.3. Procedure 163 6.3.4.4. Results 164 6.3.5. Discussion 167 6.4. Patterns of Deglottalization 168 6.4.1. Coda Deglottalization 170 6.4.2. Onset Deglottalization 172 6.4.2.1. Coast Tsimshian 173 6.4.2.2. Nisgha 174 6.4.2.3. Gitksan 176 6.4.3. Discussion 178 6.5. Conclusion 180 Chapter 7: Conclusion 181 7.1. Summary of Findings 181 7.2. Future Research 182 7.2.1. Data Collection and Analysis 182 7.2.2. Comparative Research 183 7.2.3. Experimentation 184 7.2.4. Modeling and Learning Algorithms 185 References 187 Appendix A: Gitksan Orthography 202 Appendix B: Gitksan Roots 204

iv Acknowledgments

First and foremost, I wish to acknowledge and thank the Gitksan community and the many individuals who contributed enormously to this thesis. This includes Margaret Heit, Lonnie Hindle, Laurel Mould, Delia O’Brien, Gwen Simms, Clyde and Marlain Skulsh, Fanny Smith, Jane Smith, Alvin Weget, Holly Weget, Fern Weget, Sheila Weget, and Leiwa Weget.

I thank most of all my teachers, friends, and mentors in the language, Doreen Jensen and Barbara Sennott. The Weget and Harris families also were extremely generous with their hospitality. This community has graciously shared its language with me, and I hope this thesis can be used in helping to keep the languagetaught. I owe my committee members a world of gratitude. My supervisors Douglas Pulleyblank and Gunnar Hansson were a constant encouragement; they were always available for meetings, they read my work over and over, and they allowed me to work in new directions. Henry Davis was always behind me, throughout the process. Henry was always a great to bounce new ideas around with, and pointed out all of the ramifications my work had in other areas of linguistics. To my committee: I thank you. I’m very lucky to have had Pat Moore and Joe Stemberger as University examiners; they had many helpful comments and were very encouraging. Likewise, my external examiner Sharon Hargus provided many thoughtful comments on the thesis and asked questions about things I hadn’t previously thought about. The faculty members at Cal State Fresno who got me started in linguistics and who provided constant encouragement throughout my career deserve thanks, and include: Brian Agbayani, Ritva Laury, Yukiko Morimoto, Shigeko Okamoto, Ray Weitzman; in anthropology: Walter Dodd, Charles Ettner, Roger LaJeunesse, and John Pryor; and in philosophy: Jack Weinstein. Classmates that stuck with me even after parting ways include Carlos Fierro, James Hoang, Stefan Isaksson, and Jon Petty. The one person most responsible for my development as a linguist is my friend and mentor Chris Goiston. He and Brian Agbayani are also to thank for my personal development during my formative years.

v Over the years I have had countless wonderful interactions with the UBC linguistics faculty, past and present, some of whom I would like to acknowledge here: Martina Wiltschko, Strang Burton, Rose-Marie Déchaine, Bryan Gick, Päivi Koskinen, Felicia Lee, Lisa Matthewson, Hotze Rullman, Kimary Shahin, Joe Stemberger, and Eric Vatikiotis Bateson. Through every step of the way Edna Dharmaratne looked out for me administratively, and provided unconditional friendship; I will never thank her enough. Anne-Marie Comte has always been encouraging and hospitable, and deserves a special thanks here! Auntie Svea Thompson also deserves thanks for putting up with me,

feeding me, and entertaining me while I did fieldwork. Thanks Auntie! The graduate students and postdocs at UBC were great to be around. Thanks to Leora Bar-El, Seth Cable, Mario Chávez-Peón, Carrie Gillon, Peter Jacobs, Susie Jones, Michelle Kalmar, Calisto Mudzingwa, Beth Rogers, Sonja Thoma, and Ian Wilson. A special thanks go to Karsten Koch, James Thompson, and Tyler Peterson, who have been like brothers to

me, and who will continue to be in my family after 1leave Vancouver. There are several linguists at other universities who I’ve had meaningful conversations with, who have read drafts or commented on this work, taught incredible classes I’ve been lucky to take, or who have generally inspired me: Eric Bakoviá, Juliette Blevins, Derek Chan, Donna Gerdts, Larry Hyman, Joe Pater, Sharon Rose, Donca Steriade, and Cohn Wilson. There are also many scholars in the Tsimshianic world who have helped me at one point or another. Tonya Stebbins and Fumiko Sasama have been more than generous in supplying me with published works on Coast Tsimshian. Lonnie Hindle graciously gave me a copy of the Gitksan dictionary. Bruce Rigsby has been a constant source of support and inspiration, and he has also been an incredible resource for testing ideas, digging up old papers, field slip notes, etc. Jay Powell has given me a wealth of materials on Gitksan, including the entire set of textbooks that he wrote, as well as his entire collection of

fieldnotes (which I hope to incorporate into the database in the near future). I owe all of these individuals a debt of gratitude.

I am also grateful for two Jacobs Fund grants (administered through the Whatcom Museum Society) and a Phillips Fund for Native Research Grant (admininstered by the American Philosophical Society)that have supported my field research.

vi I wish to acknowledge my parents, Dan & Chris Brown, who have provided constant support throughout all of my schooling, and who have encouraged me every step of the way. The same goes for my brother, Brandon, and the rest of my family, including the Brown, Madrigal, Sanchez, and Hernandez families. Finally, all the thanks in the world go to my wife Jennifer, and my daughters Athena and Lillian. They have given me the support that I needed to see this program through to the end, and they have given me the love that others only wish for. Thank you all so much; I am so very glad that you each of you has been beside me while I wrote this, and I am equally glad that you will be beside me for whatever adventure we choose next.

vii Dedication

For Jennzfer, my love.

viii Chapter 1: Introduction

1.1. Introduction This thesis deals with the phonology of Gitksan, an Interior Tsimshianic language spoken in northern British Columbia, Canada. The claim of this thesis is that Gitksan exhibits several gradient phonological restrictions that hold over the lexicon. In particular, the restrictions are on what consonants can cooccur within a root, and are reminiscent of gradient restrictions in other languages such as Arabic, English, and Muna. For instance, there is a prohibition on pairs of coronal consonants cooccurring, but this restriction is gradient, as there are numerous exceptions. Likewise, there are gradient restrictions on labial pairs cooccurring, and dorsal pairs cooccurring. It is claimed that these restrictions are due to a generalized OCP (Obligatory Contour Principle) effect in the grammar, and that this effect can be relativized to subsidiary features, such as place, manner, etc. For example, in addition to the gradient homorganicity restriction, there is a restriction on pairs of coronal sonorants, and a restriction on pairs of coronal obstruents. Furthermore, within the coronal obstruents, there is a restriction against pairs that agree in continuancy. Following Coetzee & Pater (2008), it is argued that these types of gradient OCP effects are best analyzed with the system of weighted constraints employed in Harmonic Grammar (Legendre et al. 1990, Smolensky & Legendre 2006), and that the phonological grammar of the language is a reflection of the structure of the lexicon. It is further argued that the way the grammar comes to reflect the lexicon is through the process of learning, and the particular learning algorithm adopted is the Gradual Learning Algorithm (Boersma 1998). Through gradual learning (i.e. lexical item by lexical item), the grammar is adjusted so that it slowly captures the patterns of the lexicon. While this type of general dissimilatory effect is not uncommon across languages, it is also claimed that Gitksan exhibits a gradient assimilatory effect among specific consonants. This type of effect is extremely rare, and is unexpected given the general conditions of dissimilation. There are two basic patterns: one is the frequency of both pulmonic-pulmonic and ejective-ejective pairs of consonants, which occur at rates higher

1 than expected by chance. The other is the occurrence of uvular-uvular and velar-velar pairs of consonants, which also occur at rates higher than chance. It is claimed in this thesis that what these patterns amount to are two types of gradient consonant harmony: a gradient laryngeal harmony, and a gradient uvularity harmony. These assimilatory patterns are analyzed using the Agreement by Correspondence approach (Hansson 2001, Rose & Walker 2004), which mandates that output correspondents agree for some phonological feature. The constraints responsible for output correspondence and correspondence-driven agreement are in competition with constraints demanding featural identity between input and output. It is claimed that just as with the OCP effects, gradual learning is responsible for shaping the grammar in the image of the lexicon. The general discussion of assimilation and dissimilation can be continued in areas outside of the root, such as reduplication. It has been claimed by Dunn (1979a) that there are robust gradient and categorical patterns of assimilation and dissimilation in Coast Tsimshian reduplicative contexts. By examining the attested reduplicants in Gitksan, it is shown in this thesis that Dunn’s generalizations do not hold true in this language. Experimental evidence, specifically a nonce-probe task, supports this claim. Results from this task suggest that speakers have an internalized knowledge of the reduplicative patterns of their language. Finally, the gradient patterns that are discussed throughout the thesis can be compared with categorical patterns of reduplication, where ejectives and glottalized sonorants deglottalize in reduplicants. An investigation of reduplication in the Tsimshianic languages yields a typology of deglottalization, and it is claimed that this typology indicates that glottalized sonorants are more marked than ejectives. There are many compelling reasons for engaging in the kind of research and focusing on the particular topics in this dissertation. The relationship that the lexicon has with the phonological grammar has become the focus of much recent research. In particular, the issue of gradient phonotactic patterns, and how to account for them has sparked much interest among phonologists. There have been numerous recent accounts of gradient phonotactics in several languages, including analyses of English (Berkley 2000), Arabic (Frisch et al. 2004), and Muna (Coetzee & Pater 2008), and many approaches in different theoretical frameworks. The issue of morphology and gradience in matching affixes to stems (the issue dealt with in

2 chapter 6) also has been of recent interest (see especially Albright 2000 on infixation in Lakhota). By addressing these particular issues in this thesis, several tasks have been accomplished. The first is documenting sound patterns that have previously been unreported. This is almost certainly due to the fact that technologies have advanced considerably over the years, allowing an electronic database to be used in this study, whereas previous researchers were forced to isolate patterns identifiable only by the naked eye. The sound patterns that are documented here also contribute to theoretical discussions. For instance, the issues surrounding the gradient consonant harmonies that are documented in this thesis will hopefully open up new discussions regarding how to account for consonantal agreement, as well as how to account for gradient patterns. Another question that will be raised, based on the overall theme of the thesis is: where does one draw the line between lexicon and phonology? Are these patterns phonological? Are they to be accounted for as part of the lexicon only? Or are they simply an accident? The study of Gitksan phonology is an important undertaking, for various reasons. The first and foremost is that like virtually every other language of the Pacific Northwest Coast, Gitksan is a highly endangered language. Because of this, it is imperative that accurate descriptive generalizations be made about the phonological system of the language spoken in the present day. There are also language-internal reasons for studying the phonology of this particular language. Gitksan exhibits several typologically rare or little- understood sound patterns which are interwoven with other aspects of the grammar, oftentimes making analysis difficult. There are, however, some very robust generalizations to be made about the phonology, once the gradient lexical patterns are taken into account. An overview of the language is provided in section 2. This includes an overview of the language area, a discussion of Interior Tsimshianic and the entire Tsimshianic language family, the history of research on the language and attempts to document it, the Gitksan orthography, and the Gitksan Lexical Database that is used as the basis for this thesis. An overview of the theoretical issues in this thesis is provided in section 3. The overall structure of the thesis, with detailed chapter breakdowns, is presented in section 4.

3 1.2. Overview of Gitksan

Gitksan is a language of the Tsimshianic family, along with Coast Tsimshian, Southern Tsimshian, and Nisgha. The Gitksan refer to their language as Sim ‘algax, which means ‘the real or true language’. However, this term is also used by Coast Tsimshian and Nisgha speakers to refer to their languages. This has created some confusion, since many published sources on Coast Tsimshian simply refer to the language as Sm ‘algyax. The language has been referred to as Gitxsan or Gitksan by scholars, or Gitxsanimx or Giixsanimax by native speakers when distinguishing it from Nisgha (Nisga ‘amx)or Coast Tsimshian (Ts ‘imsanim). Further discussion of the use of these terms can be found in Rigsby (1987).

1.2.1. Language Area

Gitksan is spoken in areas primarily along the Skeena River in northern British Columbia. The Gitksan reside also along the Bulkley and Nass Rivers. Gitksan literally means ‘people of the Skeena River’ (git- ‘people of, xsan ‘(to) gamble’; ‘Skeena River’). The language can be divided into two major dialect areas: Eastern Gitksan and Western Gitksan. The Eastern dialect, or Gaanimx (‘upstream language variety’), is spoken primarily in the villages of Kispiox (Ansbahyaxw’, ‘hiding place’), Glen Vowell (Sigit’ox, name of the nearby mountain), and Hazelton (Gitan ‘maaxs, ‘people of the place of torchlight fishing’). The Western dialect, or Geets ‘imx(“downstream language variety’), is spoken in the villages of

Kitwanga (Gitwingax ‘people of the place of rabbits’), Kitwancool (Gitwinhlguu ‘1‘people of the narrow place’), and Kitseguecla (Gigyukwhla ‘people of Jigyukwhla’, which is the name of a nearby mountain). Variation between the two dialects is minimal, and consists mostly in phonological shifts and lexical differences. There are also two other main villages in Gitksan territory that are no longer inhabited, but remain part of the Gitksan territory, and are occasionally used for summer retreats. These are Kisgegas (Gisgaga ‘as ‘people of the place of small white gulls’) and Kuldo (or Gitgaldoo ‘o,“people of the wilderness or backwoods”). The data for this thesis is based primarily on Eastern Gitksan, as my fieldwork in the language area took place in the villages of Hazelton and Kispiox.

Bruce Rigsby (personal communication) notes that the -hy- in the spelling of the place name Ansbahyaxw probably preserves an older pronunciation of the verb “(to) hide”, which is now yaxw in its singular form. The older form can be reconstructed as *xaxw

4 1.2.2. The Tsimshianic Language Family

Coast Tsimshian, Southern Tsimshian, Nisgha and Gitksan comprise the Tsimshianic language family. Coast Tsimshian and Southern Tsimshian are more distantly related to Gitksan than Nisgha is to Gitksan. These relationships can be seen in (2) below.

(1) Tsimshianic familial representations (from Rigsby 1986:25)

Tsimshianic

Maritime Tsimshianic Interior Tsimshianic

Coast Tsimshian Southern Tsimshian Nisgha Gitksan

Dunn (1979b) has noted that while the lexical differences between Gitksan (and Nisgha) and Southern Tsimshian are substantial, their phonetic inventories are near identical. This has led some researchers such as Dunn & Hays (1983) to consider the language area as more of a “dialect chain bent into a circle” than a strict linear dialect 2chain Within each respective language area there also exists. some variation across villages. In particular, Tarpent (1987) discusses several aspects of the Kincolith dialect of Nisgha, pointing out several differences such as the shift from [n] to [1], and the production of ejectives that is so weak that they are nearly 3voiced This same phenomenon was noticed and discussed to a lesser extent by Boas (1911).. Another possible noteworthy dialect area is around Kitselas Canyon; some speakers of Gitkan consider speakers from this area to constitute their own dialect group, and not one of either Gitksan or Coast Tsimshian. There are also several documented differences between Nisgha and Gitksan, as well as differences among Gitksan dialects. Rigsby & Kari (1987) note several of these differences. For instance, one difference is that while Gitksan exhibits palatal, velar and labiovelar fricatives,

2 Dunn & Hays (1983) use a slightly unorthodox method in their dialectological study, using the conservative index (CI) to determine the relationships between the various Tsimshianic languages and dialects. They reach the conclusion that Gitksan and Southern Tsimshian share many characteristics, and warn that Coast Tsimshian should not be considered a geographic/historical intermediary between the two. Nisgha also generally exhibits the same allophonic prevocalic voicing of plosives that is found in Gitksan (cf. Tarpent 1987). Voicing in Gitksan is discussed in Chapter 4 of this thesis.

5 in some Nisgha cognates there are corresponding stops. There are also differences in the spirantization of labiovelar stop 7kw! before obstruents, different environments for spirantization of post-tonic uvulars, and the existence of shortening and raising rules for underlying long vowels in Gitksan. In addition to these phonological differences, there are also many other morphological and syntactic differences. Specific details of these differences, and the sound changes responsible for them, can be found in Appendix I of Rigsby & Kari (1987). The Tsimshianic languages are bordered by Athabaskan languages in the East (Witsuwit’en, Babine, and Sekani), as well as in the North (Tsetsaut, which is no longer spoken). Immediately northwest of the region is the Tlingit language area, the Tongass dialect being geographically the closest to the Tsimshianic languages, and most relevant in terms of potentially borrowed features. Further to the north are the Tahltan (Athabaskan) speakers. To the south lie Wakashan languages, the most immediate of which is Haisla, spoken on the coast. Finally, Haida is spoken on the Queen Charlotte Islands, off the mainland coast of British Columbia. There is some evidence that linguistic borrowing may have taken place in some of these contact situations. For instance, Leer (1978) has suggested that there has been phonological borrowing due to contact between Coast Tsimshian and Tlingit. Leer discusses the similarities in “sustained” and “fading” nuclei in the Tongass dialect of Tlingit and Coast Tsimshian, though leaves open the question of which direction the influence spread. There is also a good deal of language contact in the east with Witsuwit’en. According to Rigsby (1986), there is bilingualism between the Gitksan and Witsuwit’en, and there has reportedly been linguistic influence, as well as cultural 4influence through contact. Rigsby & Kari (1987:60-68) note that while they have only isolated one verb root borrowed from Gitksan into Witsuwit’en, there are large , a number of nominals that have been borrowed in both directions (with the majority being borrowed from Gitksan to Witsuwit’en). Rigsby & Kari list several place names, plant and animal terms, terms related to material culture, and terms for social organization (see also Hargus 2007 on Witsuwit’en borrowings from Gitksan).

“In his report for the Delgamuukw v. the Queen trial, Daly (2005:205) notes that “Viewed as a single social phenomenon composed of two linguistic and historical traditions, Gitksan-Witsuwit’en society is a composite, or a blend of both the decentralized and egalitarian social values of the inland Athapaskan lifestyle and the more hierarchical and stratified social values of the north coast.”

6 This high amount of borrowing may stem from the fact that historically there was bilingualism between the two groups (Jenness 1943). Given the status of the Pacific Northwest as a Sprachbund area that is characterized by a higher than normal amount of contact phenomena and borrowing (cf. Sapir 1915, Thomason & Kaufman 1988), it would not be surprising if there were other sources for many aspects of Gitksan grammar aside from Witsuwit’en. While the Tsimshianic language family is considered by some to be an isolate, others have considered it to be a member of a much larger stock. In particular, Sapir (1921) classified Tsimshianic as a part of the Penutian stock. More recently, DeLancey et a!. (1988) and Tarpent (1996, 1997) have re-argued this point, suggesting that Tsimshianic is indeed a Penutian language. DeLancey et al. point to morphological, phonemic, and lexical correspondences to make this claim. However, the major problems that have been expressed in the literature about the Penutian stock, as well as the problems with Tsimshianic as a part of that stock (see the discussion in Rigsby 1986, as well as Campbell 1997 for a general overview) cast doubt on this relationship. I take the conservative position that the Tsimshianic family is an isolate, and not related to the Penutian stock, or any other hypothesized stocks, though nothing in this dissertation hinges on that position.

1.2.3. Interior Tsimshianic

Gitksan is closely related to Nisgha (Nisga’a), which is spoken on the lower Nass River, just north of the Gitksan area. The two languages comprise the Interior branch of the Tsimshianic language family. Some researchers consider Nisgha and Gitksan to be dialects rather than languages. For instance, Boas (1911) considered Nisgha and Gitksan to be dialects of a single language, as does Rigsby (1975), who earlier employed the term “Nass-Gitksan”. While there is a large degree of mutual intelligibility between the two, Rigsby (1987) and Rigsby & Kari (1987) provide several reasons for considering Nisgha and Gitksan as separate languages. Rigsby & Kari (1987) note that

“There is then, a broad Gitksancommunitywhich is defined in part, by the people’s own belief and public statementsthat they are Gitksanand that they speaktheir own distinct language. They are not Nisgha or Tsimshian,who each have their own native language. This assertionof Gitksanlinguisticidentityis supportedby certain

7 characteristicsof speech which set it apart from that spoken by Nisgha people.” (1987:18)

Furthermore, Rigsby & Kari point to the fact that the oral and written standards that are developing for the respective languages differ from each other in significant ways.

The exact designation does not matter for the purpose of this thesis; thus, I will follow the designations of Rigsby (1987). When critical differences arise between Nisgha and Gitksan, they will be pointed out. The differences between the two languages are extremely interesting, and important for some of the claims made in this thesis. Thus, when relevant, Nisgha forms will be presented and labeled as such. Gitksan is a seriously endangered language. According to Ethnologue (Gordon 2005), there were only 400 native speakers of Gitksan in 1999. This included a population of 220 speakers of the Western dialect, and 180 speakers of the Eastern dialect. In her extensive language survey of Gitksan, Harris (1999) cites a much higher number, listing the fluent speakers at 789 out of a total sample of 2346 individuals. While this is a seemingly rather large discrepancy, the numbers in themselves remain quite small either way, and in no way indicate that Gitksan is not endangered. Having done firsthand fieldwork in the area, I estimate the present-day number of fluent speakers to be less than the numbers cited in Ethnologue.

1.2.4. History and Documentation

There are some missionary texts that have been written in Gitksan during the late 1800s and early 1900s. A prime example is Ridley’s 1881 selection of prayers written in Gitksan (Ayer 1881). An extensive comparative vocabulary of the languages of the Pacific Northwest Coast was collected by Tolmie & Dawson (1884), who list the Gitksan dialects as part of the “Tsimshian” 5language. During the 1920s Harlan Smith, an archaeologist, documented the Gitksan knowledge of plant and animal species in the area (see Smith et al. 1997). This documentation resulted in a significant collection of ethnobotanical terms, and remains an important work in that area.

Though Tolmie only lists two dialects of “Tsimshian” in the vocabulary, presumably from Coast Tsimshian proper.

8 Perhaps the earliest linguistic work done on Tsimshianic was by Franz Boas in the early twentieth century. Boas wrote extensively on the Tsimshianic languages publishing a grammar (1911), a collection texts (1902), and mythologies6 (1916). Much early of ethnological work was also conducted by Marius Barbeau ,and William Beynon. Specifically, relevant material can be found in Anderson & Halpin (2000). Later work on the Gitksan language includes the work of Bruce Rigsby, who has published several articles on the language, as well as an unpublished grammar (1986), and is generally regarded as the foremost expert on the language. Pedagogical materials have been produced by Jay Powell, Vickie Jensen, and Russell Stevens (Powell & Stevens 1977, Jensen & Powell 1979-1980). These include language textbooks for children, as well as language textbooks for the adult curriculum. Other teaching materials have been compiled by Jay Powell and Bruce Rigsby and used in various contexts. A practical dictionary was compiled by Lonnie Hindle, a native Gitksan speaker from Hazelton, and Bruce Rigsby (Hindle & Rigsby 1973). This dictionary is especially important, as it remains the only published and available lexicon of the language. There are also numerous language teachers in the area who have independently developed their own pedagogical materials/curriculum, and continue to do so, such as Delia O’Brien, Jane Smith, and Fern Weget.

1.2.5 The Gitksan Database The present work is based in large part on an electronic database of Gitksan morphemes, words, and phrases. This database was developed using Microsoft Access, and contains 1601 entries. The database subsumes the Hindle & Rigsby (1973) dictionary, portions of Rigsby (1986), as well as material from my own field notes from 2004-2008. Each entry is coded for the source of the data. This database is the foundation for a dictionary of Eastern

6 Interestingly, Bruce Rigsby (personal communication) has found no evidence that Franz Boas ever worked on Gitksan, but rather only worked on Coast Tsimshian and Nisgha. This conclusion was drawn based on careful study of Boas’ published works, as well a study of Boas’ correspondence at the American Philosophical Society in Philadelphia in 1969.

9 7Gitksan. The list of roots drawn from the database for the purpose of this study is presented in Appendix B. The entries in this database are morphologically broken down into roots, words, clitics, and compounds. At present, there are 645 roots: 306 are verbal roots, 308 nominal, and 31 that for now are listed as “other” (such as adverbs, wh- words, etc.). For the purposes of counting roots, suppletive pairs of words, as are sometimes found in singular/plural forms, are counted as separate roots, primarily because they do not share a relationship phonologically. Each entry contains a morphological segmentation, and in cases where there does not appear to exist a free root, the apparent morphology has been stripped and an entry for a bound root has been provided. This database provides the basis for Chapters 2 and 3, where static patterns over the

Gitksan lexicon are discussed. It also provides the data for Chapters 4 and 5, where additional lexical patterns involving laryngeal features and dorsal consonants are laid out, and Chapter 6, where patterns involving reduplication are dealt with. In addition to these broad lexical searches, the database has provided additional forms where needed. It is a long-term goal to extend the content of this database beyond what is presented in this thesis so that further studies can be performed with an increasing degree of accuracy. The data presented in this thesis come from the database; thus, for the most part, many words or morphemes were originally found in Hindle & Rigsby (1979) or Rigsby (1986), and then were re-elicited. Where consultants did not recognize a particular word, it was noted that the word was archaic or dialectal.

1.2.6. Orthography

To ensure accessibility for speakers and teachers of the language, the practical orthography developed by Hindle & Rigsby (1973) is used in the database. This orthographic representation was converted to single characters (i.e. digraphs and trigraphs were eliminated) for the purpose of doing phonological analyses. The Hindle & Rigsby orthography is presented in Appendix A.

The project outlined here is not simply an index of words in the language, but rather, is a full dictionary that includes words/morphemes and their meanings.

10 1.3. Theoretical Overview

The purpose of this dissertation is twofold: it attempts to provide an accurate phonological descriptive account of the Gitksan language for use by scholars and students of the language. It also serves to situate the phonological structure of the language into a much broader theoretical and typological picture. It will be shown that the sound patterns exhibited by Gitksan are oftentimes typologically interesting, and sometimes exceptional; and that these patterns will help us in theory construction, since theories of this type (i.e. typological) by defmition aren’t based on the analysis of one phonological system.

1.3.1. Documenation and Theory

In a recent article on endangered sound patterns, Juliette Blevins (2007) offers several practical reasons for documenting endangered languages and their sound patterns. One reason that Blevins gives is that language documentation can play a central role in language revitalization and maintenance. Another is that documenting endangered languages can reveal rare or unattested sound patterns, or that new examples of attested patterns may be uncovered. This thesis is done in the spirit of Blevins’ suggestions. On the one hand, this work is being undertaken in order to contribute to a detailed phonological and morphological description of Gitksan, such that the content of the thesis can be extracted and developed into materials which can be exploited for pedagogical uses. This will also provide an accurate description of the language such that scholars can extract the linguistic generalizations. Related to this is the construction of the database used in this study, which is ultimately being put together for future use as a dictionary of Eastern Gitksan, and will serve as a tool in comparative Tsimshianic research. On the other hand, this thesis is written as a theoretical contribution— one that attempts to place Gitksan into a theoretical background, and to illustrate the typologically interesting and rare sound patterns, as well as the patterns that are more widely attested. In much the same fashion this thesis attempts a methodological contribution. Aside from a few cases, for the most part endangered languages, and especially the native languages of British Columbia, have historically received less attention in theoretical linguistics than more well-studied languages. In recent years, however, there has been growing interest in what these languages have to offer to linguistic theory. There has

11 also been a growing interest in using new experimental methods in order to conduct this research. Some of the methods described in this thesis are not conventional among studies of First Nations languages. These include large-scale investigations of lexical patterns, and nonce-probe tasks. It is a goal, then to encourage this type of approach in the study of these languages.

1.3.2. Constraint-Based Theories and Gradient Phenomena It is also helpful to turn to the specific nature of the theoretical setting of the thesis. Many sections rely heavily on discussions of Optimality Theory (OT; Prince & Smolensky 1993 [20041). Within OT, there are families of constraints which govern the output of phonological forms. Candidate forms are generated, and these candidates may violate constraints. If two candidates, A and B are in competition, and A violates a higher-ranked constraint that B does not, then A is eliminated from the competition. B can incur any number of violations of lower ranked constraints, but it will not change the outcome. In this way, lower ranked constraints cannot overpower a higher ranked constraint. While OT has become orthodox in the field of phonology at the moment, the theoretical core of this thesis is based on Optimality Theory’s cousin, Harmonic Grammar (HG; Legendre et al. 1990, Smolensky & Legendre 2006). While HG shares with OT the basic idea that the mapping from input to output is done in parallel, it diverges from it in replacing the notion of strictly ranked constraints with the mechanism of weighted constraints. The idea is that the more highly a constraint is weighted, the more severe the penalty for violating it. This architecture differs from standard OT in that it allows for ‘gang effects’, whereby multiple violations of a lower-weighted constraint can ‘gang up’ on a higher-weighted constraint, and thus force a different candidate to be selected. This type of effect is impossible in standard OT, where a lower-ranked constraint can never overpower a higher-ranked constraint to determine a winning candidate. Another major difference between the two theories is in how they account for gradient versus categorical phenomena. In standard OT, all phenomena are “categorical” in the sense that a single output is selected consistently for a given input. However, it has been shown that because constraints are only ranked by a domination hierarchy, gradient lexical

12 patterns do not receive a straightforward analysis in OT (Berkley 2000, Frisch et al. 2004). By ‘gradient lexical pattern’, what is meant is a generalization over the lexicon that is not absolute. A prime example is the type of OCP effects mentioned in section 1.1: in many languages, there is a tendency for pairs of homorganic consonants to be avoided in roots, but the exceptional (and non-exceptional) lexical items surface consistently every time (i.e. there is no “variation”). This generalization is gradient, as there are exceptions to it. In OT a hierarchy of ranked, strictly dominated constraints can derive a consistent output at each evaluation, but there is no mechanism that can capture the gradience of the pattern itself. While standard OT has no real way of accounting for a gradient pattern like this, the system of weighted constraints in HG does. The HG architecture allows for the modeling of gradient well-formedness, whereby the summed violations of constraints for a particular candidate add up to a harmony score, and the harmony scores between candidates can be compared, yielding an acceptability score. This acceptability score is in theory directly correlated with the degree of attestation of a given pair of consonants in the lexicon. It is for this reason that HG is adopted as the theoretical framework of this dissertation. The exact details of these architectures, including the crucial differences between strict domination in OT and weighted constraints in HG will be dealt with in chapter 3.

1.4. Structure of the Thesis This thesis is structured as follows: Chapter 2 outlines the patterns of consonantal cooccurrence that are found within roots across the Gitksan lexicon. In particular, the phonological shape of roots, and the gradient consonant cooccurrence patterns found in roots is discussed. It will be shown that there is a gradient OCP[placej restriction in the language, similar to what has been found in studies of Arabic, English, and Muna. These restrictions are modeled by means of the weighted constraints of Harmonic Grammar in chapter 3. A system of weighted constraints is outlined, and a learning algorithm is adopted which correctly models the structure of the lexicon.

The laryngeal features are discussed in chapter 4. Picking up where chapter 3 leaves off, a gradient glottalization agreement pattern is identified in the Gitksan lexicon, and this pattern is accounted for with (weighted) correspondence constraints that enforce featural

13 agreement between consonants (Hansson 2001, Rose & Walker 2004). The phonological process of plosive voicing is also discussed in this chapter. The status of voiced obstruents is explored by means of consonant cooccurrence restrictions. It is shown that while voicing is allophonic and not underlying, cooccurrence restrictions are still sensitive to whether a consonant surfaces as voiced or not.

Chapter 5 deals with a phenomenon of gradient uvularity agreement among dorsals that is touched upon in chapter 2. Like the case of gradient glottalization agreement, uvularity agreement is accounted for by using correspondence constraints. Crucial to the analysis is the class of “guttural” consonants in the language. These include the laryngeal and uvular consonants. The principle process that unites these sounds as a class is vowel lowering. This process is observed in morphological contexts, and the representations provide the basis for the correspondence constraint definitions. Chapter 6 investigates reduplicative phenomena. There are two main aspects that are focused on: gradient patterns between reduplicative template shape and properties of the base, and the deglottalization of consonants in reduplicants. It is claimed that there are no gradient (or categorical) generalizations to be made concerning template shape and base properties. It is also claimed that reduplicative deglottalization across the Tsimshianic languages fills out a typology whereby positional faithfulness constraints allow glottalization to surface in privileged positions in some languages. Chapter 7 provides a brief conclusion, and discusses possible directions for future research.

14 Chapter 2: Patterns Across the Lexicon

2.1. Introduction

Like many other languages of the world, Gitksan exhibits restrictions on what types of consonants can cooccur in words. To date, there are many quantitative studies that have focused on this type of property in various languages: famously, Arabic (Greenberg 1950, McCarthy 1988, Frisch et al. 2004), but also English (Berkley 2000), Latin (Berkley 2000), French (Berkley 2000), Hebrew (Everett & Berent 1997), Muna (Coetzee & Pater 2006, 2008), JuI’hoansi (Miller-Ockhuizen 2003), Javanese (Mester 1988, Yip 1989), Japanese (Kawahara et al. 2006), Russian (Padgett 1995), Ngbaka (Thomas 1963, Mester 1988, Broe 1995, cited in Berkley 2000), Berber (Elmedlaoui 1995), Amharic and Chaha (Rose & King 2007, Rose & Walker 2004), Yucatec Maya (Noguchi 2007), several creole languages (Kinney 2005), etc. For example, according to Berkley’s observations of English, words that have pairs of homorganic consonants such as tot /tat/ or dot /datJ are fewer than expected (based on the independent frequencies of the relevant segment types in Ci and in 2C of 12CVC words) than words with heterorganic pairs of consonants, such aspot/pat! or cot /kat!. This can be contrasted with labial consonants in Arabic, which are prohibited from cooccurring in adjacent position (i.e. adjacent on the same tier; cf. McCarthy 1994). Thus,

{f, b, m} - {f, b, m} constitute categorically prohibited sequences within Arabic roots (i.e. a root like */bft/ would be impossible in the language). The restrictions that are found in the languages listed above are usually primarily on place of articulation, but the strength of the restriction is also a function of how similar two consonants are in other respects. Gitksan exhibits these types of dissimilatory place restrictions gradiently, but also exhibits different assimilatory patterns among pairs of consonants. Some of these assimilatory patterns are surprising, especially given the larger context of similarity avoidance that obtains in the language. It will be proposed in Chapter 3 that, following Coetzee & Pater (2006, to appear), these restrictions are the result of OCP[place] constraints which are relativized to consonant pairs agreeing in various relevant distinctive features, and in competition with faithfulness constraints. Furthermore, these

15 constraints are numerically weighted, which captures the effect that the lexicon has on a phonological grammar. The chapter is structured as follows: Section 2 provides an overview of the consonantal cooccurrence patterns over the Gitksan lexicon. This includes an overview and an outline of the method that was employed in searching the lexicon, as well as the statistical methods for determining significance. Results indicate that there are both dissimilatory and assimilatory patterns in the Gitksan lexicon. Section 3 provides a summary and discussion of the findings.

2.2. Consonantal Patterns Over the Gitksan Lexicon The goal of this chapter is to explore certain static distributional patterns over the Gitksan lexicon. An obvious reason for this type of research is the growing literature on lexical patterns (in part due to methodological advances which allow for large corpora to be examined; cf. Fanselow, Fery, Vogel, & Schlesewsky 2006), and the sometimes gradient nature of these patterns. A more interesting reason has to do with the consonant inventory of Gitksan, which, like Arabic, has several coronal segments, but also has a series of dorsal segments which are nearly evenly matched with the coronals. The inventory of the language is relevant to these kinds of issues since it has been claimed by Frisch et al. (2004) that phonological similarity drives OCP effects, and that similarity is a function of a language’s inventory structure. It is also possible that some patterns, in particular gradient ones, are not so easy to detect with the naked eye; thus, searching through the database for these more subtle patterns is likely to be informative about the phonology of the language (assuming that gradient distributional patterns are within the realm of the phonological component of a grammar, broadly speaking). In this section the consonant cooccurrence patterns in the Gitksan lexicon are outlined, and the methods used to determine these patterns are laid out. The results from this study are presented in 2.2.2.

2.2.1. Overview and Method

As mentioned above, the most famous and best studied case of consonant cooccurrence restrictions is Arabic (Greenberg 1950, McCarthy 1989, Pierrehumbert 1993, Padgett 1995,

16 Frisch et al. 2004). What obtains in Arabic is this: within root morphemes, pairs of homorganic consonants are severely underrepresented, while heterorganic consonants cooccur freely. There are other factors that contribute to the effect, though, such as major class and maimer. The cooccurrence restriction is stronger for coronal sonorant-sonorant pairs, coronal fricative-fricative pairs, and coronal plosive-plosive pairs than for coronal- coronal pairs in general, as seen in the table in (1) below (there is an underrepresentation of sonorant-fricative, fricative-plosive pairs, etc., though these are less strongly underrepresented). Thus, major class features and manner features contribute to the effect, as the coronal-coronal pairs overall (as opposed to coronal obstruent-obstruent, etc. pairs) are not underrepresented in the same way. Thus, the more similar a pair of consonants, the stronger the prohibition on their cooccurrence. So, for example, there are no roots in Arabic with the form /d t Cl (with a homorganic pair of consonants differing only in voicing), while there are two roots with the form /d s C/ (with a homorganic pair of consonants differing in voicing and in continuancy), and 4 roots with /d g C/ (where the pair of consonants is heterorganic) (Frisch et al. 2004:185). In order to determine whether a pair of consonants occurs with greater or lesser frequency than would be expected if all else was equal, an Observed/Expected (O/E) ratio is determined (Frisch et al. 2004; the ratio is originally due to Pierrehumbert 1993). In order to determine the OlE ratio for a pair of consonants, the number of observed occurrences of a pair of consonants is divided by the number of words that the pair of consonants would be expected to occur in if there were no OCP effect. The Observed value is simply the number of occurrences that are observed for a given consonant pair. The Expected value is slightly more complicated to generate. E is calculated by first determining what the probability of a consonant pair is, which is the product of the probabilities of each of the consonants in either 1C or 2C position. This probability, multiplied by the total number of pairs of consonants, derives the expected frequency (E). To take a concrete example, take a dorsal-coronal pair such as /k’ats/ ‘to land, arrive’. First, the frequency of lk’/ in 1C is multiplied by the frequency of/ts/ in .2C Likewise, the frequency of/ ts / in Ci is multiplied by the frequency of /k’/ in .2C These two values are then summed to derive the expected frequency of occurrence, and then multiplied by the total number of 12C/C pairs in the lexicon to derive the

17 Expected value if consonants were combined freely without any place of articulation effects.

Its-k’! and /k’-ts/ are being conflated in terms of E values since, following Frisch et al.

(2004), linear order is not being paid attention to.

An 0/F ratio that is equal to 1 indicates that the number of observed and expected occurrences of a consonant pair is the same. A number smaller than 1 indicates an underrepresentation of a consonant pair; i.e. that there are fewer occurrences than would be expected if everything else was equal. A number greater than 1 indicates overrepresentation of a consonant pair; i.e. that there are more occurrences than would be expected if everything else was equal. Take a particular ordered pair, such as [lVn]. In everything that follows, the o values for both [lVn] and [nVl] are tallied and divided by the sum of their respective E values. If the resulting value is above 1 (hypothetically if there are 50 observed occurrences and only 40 expected, OlE = 1.25), then 1-npairs are overrepresented; if it is below 1 (with hypothetical 30 observed and 40 expected occurrences, OlE = 0.75), then 1-n pairs are underrepresented. It is important to note that there is a difference between the notion of under/overrepresentation and (low/high) frequency of occurrence. This is easily illustrated with a hypothetical example: it is possible for a lexicon to contain several times as many instances of [sV+Jas it does of [mVp’], but for {sV+]to nonetheless have a much lower OlE value than [mVp’J. If, say, labials are on the whole much less frequent than coronals, then the scenario just mentioned may result. For instance, if [mVp’] pairs only occur 25 out of an expected 50 times, then their OlE value = 0.50. [sV+]pairs, on the other hand, may occur 30 times out of an expected 100 (OlE = 0.33). While [sV+]pairs are higher in frequency, they are lower in terms of OlE. It is also important to note that the degree to which a pair of consonants is under- (or over-) represented can be statistically significant or not; discussion of the determination of statistical significance for underrepresented pairs in Gitksan will be discussed later in this section. The OlE ratios for the various major places in Arabic, including subdivisions of coronal place, are listed below for adjacent pairs of consonants in roots (nonadjacent pairs of consonants yield similar, but higher results). The terms “adjacent” and “nonadjacent” are not to be taken here in a literal sense for the Arabic data; adjacency here simply refers to the

18 ______

relation between either consonants or vowels within a string (such that consonants in both CC and CVC would be adjacent, but consonants separated by another consonant, as in

C,.. .C. . .C would not be adjacent (cf. McCarthy 1979). These values will be useful in comparison with the Gitksan data that will be addressed shortly.

(1) O/E table of Arabic adjacent consonant pairs (adapted from Frisch et a!. 2004:186) Labial Dorsal Dorsal- Guttural8 Coronal Coronal Coronal b fm k g q Guttural hfh? Son. Fric. Pbs. x’ irn Oószs tdtd zc$ Labial 0.00 Dorsal 1.15 Dorsal- 1.35 0.07 Guttural Guttural 1.17 1.04 0.07 Coronal 1.18 1.48 1.39 Sonorant Coronal 1.31 1.16 1.41 1.26 1.21 Fricative Coronal 1.37 0.80 1.43 1.25 1.23 0.52 Plosive

As can be seen in the table above, there is an exceptionless prohibition on pairs of 9labial consonants in Arabic roots (OlE = 0.00). There is also an almost exceptionless prohibition on dorsal pairs (OlE = 0.02). The same holds for pairs of coronal sonorants, coronal fricatives, and coronal plosives. What is notable is that pairs combining coronal fricatives and coronal sonorants are overrepresented, as are pairs combining coronal sonorants with coronal plosives. However, pairs of coronal obstruents that disagree in continuancy are highly underrepresented, as the plosive-fricative pairs have an OlE value of 0.52, a value that is much higher than that of the labial and dorsal pairs, but lower than for any of the other ‘mixed’ pairs. The significance of this is that this underrepresentation indicates that the restriction is on pairs of coronal obstruents, and not just on pairs of coronals. Finally, coronal fricative and coronal plosive pairs are each extremely underrepresented, with OlE values of

8 The class names and order are slightly different here than in the original context; this is partly to keep the coronals linearly together for comparison, and also to highlight the uvular fricatives as belonging to the Dorsal- Guttural class. Although [w] is classed as labial, it does not appear in the corresponding table found in Frisch et al. (2004:186).

19 0.04 and 0.14, respectively. What these numbers indicate is a gradient sensitivity to similarity. In other words, within the coronals, [cor, -son, ctcont] segments are the most severely restricted, those that are [cor, ason] but differ in [±cont] (as in the coronal- fricative/coronal plosive cell in 1) are next most severely restricted, and finally, segment pairs that only share [con (and differ in both [±son] and [±cont]) are the least restricted. It is important to note that these patterns are not specific to Arabic or (as will be shown) Gitksan. Rather, similar patterns have been attested in English, Latin, French (Berkley 2000), Muna (Coetzee & Pater 2006, 2008), Russian (Padgett 1995), and many other languages. In other words, the pattens are cross-linguistically robust, and are not specific to either Arabic or Gitksan. Given these figures from Arabic, it is now relevant to consider how consonants are restricted in the lexicon of Gitksan. The consonantal inventory of Gitksan is given in (2).

(2) Gitksan consonant inventory Labial Coronal Velar Uvular Glottal plosive p t ts kkW q ejective p’ t’ ts’ t+ k’ k” q’ fricative s x x”' x h

sonorant m w n 1 j glottalized 10sonorant m’ w’ n’ 1’ j’

There are several non-trivial assumptions behind the structure of this inventory. For example, the glides 1w,w’/ have been placed in the labial category, though articulatorily they are labio-dorsals. This is parallel to treatments of /w/ in Arabic (Frisch et al. 2004), English (Frisch 1996), and Muna (Coetzee & Pater 2006, 2008). Cooccurrence restrictions also indicate that there is a gradient restriction between 1w w’/ and labials (OlE = 0.33), but not between /w w’/ and dorsals, which have an OlE below 1 (OlE = 0.80), but is not significant.

This suggests that with respect to place cooccurrence restrictions, /w w’/ belong to the labial class, and not the dorsal class. Similarly, the labialized obstruents / k”’ k”” XW/ have been

10 The glottalized sonorants will be indicated with a following apostrophe; this is not intended to reflect their phonetic timing properties.

20 classed under the velar place of articulation due to a cooccurrence restricton between these sounds and non-labialized dorsals (OlE = 0.43), though clearly there is a secondary labial gesture associated with these segments that contrasts them with their plain counterparts. Since their primary place is velar, however, they have been put in this category. This issue will resurface when labial-velar cooccurrences are discussed. The categorization of the glides /j,j’l is also somewhat controversial. These segments do not exhibit cooccurrence restrictions with either the coronals (OlE = .98, n.s.) or the dorsals (OlE = 0.89, n.s.). Furthermore, the actual featural representation of palatal consonants is up for debate; therefore, the palatals will simply be considered as specified for coronal place here (cf.

Clements & Hume 1995). This is in contrast to Frisch et al.’s analysis of Arabic, where UIis treated as a dorsal”. The identification of the occurrence of consonant pairs was performed by extracting the entire set of roots from the Gitksan database (discussed in chapter 1). The total number of roots in the database amounted to 645 out of 1601 total entries. This included 308 nominal roots, 306 verbal roots, and 31 that were classified as “other”.

(3) Statistical breakdown of roots in database (out of 645): 308 nominal roots 47.8% 306 verbal roots 47.4% 31 “other” roots 4.8%

Many roots are “category-flexible”, meaning that they can be used as either nouns or verbs

(though I am not making any claims about “category-neutrality” in the theoretical sense, as discussed, for instance, in Sasama 2001 for Coast Tsimshian). There are challenges posed both by the database and by the phonotactics of Gitksan. These challenges will be addressed below. Many entries in the database that were potentially classified as free roots, but which appeared to have what might be frozen morphological endings (such as the voice suffix x’”) were set aside and not included in these calculations, so the total number of roots used in this

There are actually some good concrete reasons for considering [j] and U’]as dorsal rather than coronal in Gitksan. For instance, the velar consonants are strongly “fronted”, and nearly palatal in some contexts, much like many other languanges of the Pacific Northwest. Thanks to Pat Shaw for pointing this out to me.

21 study is conservative. That being the case, this has resulted in a relatively low number of items that are being studied. Compared to studies of English such as in Berkley (2000), which included words from the MRC Psycholinguistic 12Database (1987), and Dmitrieva & Anttila (2008) who observed CVC syllables in both the CELEX database (Baayen et al. 1995) and the CMU Pronouncing Dictionary’ or studies of Arabic such as Frisch et al. (2004) which included 2674 verbal3 roots from Cowan’s (1979) dictionary, or Coetzee & Pater’s (2006) study of 5854 Muna ,roots from van den Berg & Sida’s (1996) dictionary, the 645 roots under study here is small. Since there is a chance that OlE values may be misleading in this regard, it is necessary to calculate the significance of their deviation from 1.00, which in effect derives values for statistical significance. This was done using a Monte Carlo procedure (Kessler 2001, Martin 2005, 2007; the use of the procedure, and the presentation, owe much to Martin 2007)’. A Monte Carlo procedure is simply an algorithm which involves an element of simulating “random chance”; for this study, it is used as a method of determining how likely it is that the deviation of a value from 1.00 is significant, rather than just a product of random chance. Taking all instances of CVC roots as an example, in order to determine whether the deviation is significant, first the CXVCYpairs among all 12CVC pairs had to be counted (where C, C are the relevant segment types). For this “toy” example presented for illustrative purposes, the first 10 CVC roots in the database (arranged by English gloss) were selected, and place of articulation is being observed for the 1C and 2C positions with the goal of identifying coronal-coronal pairs.

12 The MRC Psycholinguistic Database is a machine-searchable dictionary. Because there were several different parameters that were explored in Berkley (2000), I am uncertain as to the total number of forms that were used in the study. Suffice it to say that this particular English corpus, and others used in similar studies, such as the CELEX database are massively larger than the Gitksan corpus under investigation here. 13 Available at: ftp://ftp.cs.cmu.edulafs/cs.cmu.eduldatalanonftp/project/fgdataldict/ 14 In order to run the Monte Carlo procedure, the list of Gitksan roots was first modified such that all allophonic alternations were “undone” and underlying forms were used. This was done in Pen with a script written by Gunnar Hansson. Since the Gitksan orthography employs di- and trigraphs, these were converted into single symbols, which facilitated the extraction of all C-C pairs. This extraction procedure allowed for lists of cooccurring consonant pairs to be generated. These lists were then used as input to the statistical method by means of the R software package; scripts for performing the Monte Carlo procedure for each relevant pair of consonant classes were also written by Gunnar Hansson. Thanks to Gunnar Hansson for these scripts.

22 (4) Word list for Monte Carlo procedure Word 1C 2C ?a ‘root species’ x lu:x ‘alder tree’ 1 X hon ‘anadramous fish’ h fl ?a:t ‘ashes’ t

gum ‘ashes, fly-ash’ —> g m dil ‘bag’ d I n’ax ‘bait’ x +it ‘ball’ ma:s ‘bark’ m 5 moI ‘barrel’ m I

The number of instances of a particular pair of consonants (here, coronal-coronal pairs are being observed, which is indicated by shading) is then counted to derive the Observed value. Then the list can be submitted to the Monte Carlo procedure. This involves holding the 1C column constant, while shuffling the 2C column. This is illustrated in (5), where the homorganic 1C and 2C pairs are highlighted for the initial state (i.e. the real lexicon), for the first shuffling of the 2C column, and for each subsequent iteration of the 15procedure. This shuffling procedure simulates random chance.

15 Identical pairs of consonants tend to be an exception to OCP effects (see discussion on pg. 29). This being the case, identical pairs of consonants are being removed for these counts, and this is done at every iteration of the procedure.

23 (5) Iterations of the Monte Carlo procedure Initial State First Shuffle Second Shuffle Ci 2C Ci 2C 1C 2C ? x x x

1 x I 1 h n h h t

? t 7 m g m g — g d d n d x “ n’ m X

I t’ I I + S m S m ‘ m n m m m

Thus, for the set of all 12CVC pairs, 1C was held constant, and 2C was randomly re-shuffled a set number of times, effectively approximating random chance. The frequency of CXVCI pairs among the re-shuffled 12CVC set is then counted. The procedure is repeated n times (following Martin 2005, 2007, n = 10,000). Every one of the 10,000 shuffles creates a new 0 value (with the E value remaining constant and unchanging; this is because it is a reflection of the segmental frequencies within each of the 1C and 2C positions, which are not affected by the shuffling), where these 10,000 artificially generated 0 values will cluster around the E value. The E value can in a sense be read off of the distribution as the mean of the 10,000 0 values. Since shuffling creates random chance, a probability (i.e. chance) distribution of potential 0 values (and hence potential OlE ratios) is generated for the cooccurrence class in question (for the example above, pairs of coronal consonants). The probability distribution is then compared to the actual frequency count in the lexicon in order to determine how likely it is that the actual frequency occurrences are due to chance. The results of the Monte Carlo procedure for transvocalic coronal-coronal pairs (for the entire list of roots in the language) is presented as a histogram in (6). The x-axis represents the number of coronal pairs in both the actual lexicon, and the simulated lexica. The actual occurrences (the 0 value) are indicated with the diamond on the x-axis. This value can be located with respect to the 95%, 99%, and 99.9% confidence intervals that are

24 displayed (which translate to p-values of 0.05, 0.01, and 0.001, respectively). The y-axis represents the number of iterations in which each 0 value along the x-axis was found.

(6) Non-identical Coronal-Coronal CVC pairs

Expected (E = 33:11)

LL-.CD

0 0 C’)

C) CD .10 CN 0 CD .4-0 0 If)CD confidence intervals: z C CD Observed _l (0 = 297)

Lf) ‘rH of

[1’ E

260 300 320 340 360 360 Observed Cor-Cor counts (in 10,0(0 Monte Carlo lexica)

For this particular example, there are 297 Observed pairs of coronal consonants in the Gitksan database, with an Expected value of 331.1. 10,000 iterations of the Monte Carlo procedure resulted in the distribution in (6), where each vertical line representeds the 0 value for a set of simulated lexica. It can be seen that the 0 value for the Gitksan lexicon falls

25 within the 99.9% confidence interval, meaning the underrepresentation is statistically significant (p <0.01). One important reason why the Monte Carlo procedure works, especially with respect to “reshuffling” ,2C is that this preserves the relative frequency distribution of individual Cs in 1C and 2C position, respectively. The E value of the OlE ratio takes not only segmental frequency, but this kind of position-based segmental frequency into account.

In the database there are 1410 root .. .C. . .C... pairs. One challenge that is faced when dealing with the roots of Gitksan is the great diversity of root shapes (CVC, CCVC, CVCVC, CCVCC...). This is in stark contrast to the triliteral roots found in Arabic, and it also poses some analytical challenges. One immediate question is how to define consonantal contexts so that there is consistency in how C-C pairs are classified. This is achieved by adopting a three-way classification into adjacent pairs, transvocalic pairs, and long-distance pairs:

(7) C-C classifications Adjacent (CC): An uninterrupted consonant sequence Ex: sg”a ‘to be square’ (sg”’) gloq ‘shame’ (gl) swan ‘to blow’ (sw) ba:sx ‘to fear’ (sx) gamk ‘to be hot’ (mk)

Transvocalic (CVC): A consonant sequence interrupted by only a vowel (short or long) Ex: mii ‘bum’ (m-+) m’uk’ ‘to catch fish’ (m’-k’’) hats ‘to bite’ (h-ts) saq ‘to be sharp’ (s-q) lip ‘sew’ (l-p) t’o:q ‘to eat’ (t’-q)

Long-distance (Cx. . . C... Cr): A consonant sequence interrupted by at least one consonant (and any number of vowels)

Ex: sins ‘to be blind’ (s. . .s) w’e:q’s ‘to find’ (w’...s)

xsta: ‘to win’ (x.. .t)

26 gida ‘to ask’ (g. . .x)

q’esxq ‘to be green, unripe’ (q’. ..x), (q’. . .q), (s. ..q)

Of the 1410 total pairs, 270 were pairs of adjacent consonants not separated by a vowel (CC), 693 were transvocalic (CXVCy),and 447 were long-distance pairs, meaning they were separated by at least one other consonant (Cx...C. . .C).

(8) Breakdown of .. . C. . . C... pairs (1410 total) adjacent (CC) 270 transvocalic (CXVCY) 693

long distance (Cx...C. . .C) 447

It is worth pointing out that the morphology of Gitksan is not the same as the nonconcatenative morphology of the Semitic languages, including Arabic. Arabic roots are triconsonantal (though occasionally roots are found with two or four consonants), and these roots are inflected by inserting vowels into the segmental sequence. In an Arabic root with a triconsonantal sequence 123C-C there is essentially no meaningful difference between (using the terminology defined, here) adjacent and transvocalic pairs. This type of sequence can surface as 123CVC 123CVC CVC or even VCCiC,VC in some cases, depending on the, inflectional, form of123 the word. Since 23Arabic employs this type of nonconcatenative morphology, 1C and, 2C will sometimes be adjacent, and sometimes transvocalic. This is not the case in Gitksan, where this classification will be consistent for each particular root. It is worth noting, though, that the long-distance category corresponds more or less directly to the category of “non-adjacent” C-C pairs in Arabic (e.g. as in Frisch et al. 2004). There are, of course, some potential issues with this classification. For instance, in both CVC.CV and CV.CV, the 12C-C pair (i.e. the first pair of consonants) gets classified in the same way (as ‘transvocalic’), regardless of whether the consonants are tautosyllabic or heterosyllabic. Likewise, 1C and 3C in a sequence like CVCC are classified as being “equivalent” to 1C and 5C in a sequence such as CCVCVCC, even though the former are likely to have much more in common with 1C and 2C of CVC. When more detail is added in the future, including the positions in syllable structure that consonants occupy, it is expected

27 that an even tighter classification system can be established. The cruder classification system outlined above is used in order to keep overall numbers large enough in the hope that under overrpresentation results reach significance. The bulk of this study is concerned with the transvocalic pairs, though some consonant types were additionally observed for how they behaved in long-distance contexts. Aside from the discussion of completely identical consonant pairs and homorganic adjacent pairs, consonant clusters (i.e. their phonotactics, their relation to syllable structure) are not treated at length in this dissertation. Once all of the contexts were collected, Observed/Expected (O/E) ratios were calculated. This was done to determine how many

C. . . C, pairs occur, relative to how many ought to occur by chance, and the significance of how far these individual O/E ratios deviate from 1.00 (if at all) was assessed using the Monte Carlo procedure described above. The literature has shown that cases of complete identity are sometimes exempt from OCP[place] restrictions (MacEachern 1999, Frisch 2004, Gallagher 2008). Muna is one such example, where there is a restriction on homorganic consonants, but identical consonants cooccur freely (Coetzee & Pater 2008). In other words, while pairs of homorganic consonants may not be tolerated, pairs of consonants that are completely identical will be allowed. Since this condition is prevalent among many other languages, it is first necessary to determine whether this is also the case in Gitksan. If we start with the “adjacent” pairs of consonants (CC), we find that there is a very significant underrepresentation of identical pairs (O/E = 0.04, p < 0.00 1). This is a reflection of a ban on morpheme-internal geminates in the language. The sole exception is the root /annoq/ ‘to permit’. Since a ban on morpheme-internal geminates may be considered a separate property of the phonology, we should consider the other types of consonant pairs (transvocalic and long-distance). In order to determine what overall effect adjacent pairs have, the adjacent class was excluded from the calculation, and only the class of “non-adjacent” pairs (including transvocalic and long-distance) was observed. For the transvocalic pairs, identical consonants received an O/E value of 0.91, which is slightly underrepresented, but not significantly so. For long-distance pairs the O/E value is 0.97, and again, the underrepresentation is not significant. In other words, C. . . C, pairs are not disfavored, except under direct adjacency (which is comparable to the notion of geminate). Thus, we can conclude that the significant

28 underrepresentation that the class of adjacent pairs of identical consonants exhibits is due to other phonological constraints on geminate structures (constraints banning geminates, with one possible exception). Identical pairs behave in a predictable fashion, where they constitute an exemption to the OCP effect. Thus, the results presented from here on will be based only on non-identical pairs of consonants (unless stated otherwise).

2.2.2. Results

The results of this investigation into the patterns of consonant cooccurrence in Gitksan are presented below. First to be discussed are the gradient dissimilatory effects; that is, the OCP type of effects that have been found in previous studies of Arabic, English, etc. where consonants are restricted in their cooccurrence. Next, some gradient assimiliatory effects will be discussed. These effects are somewhat novel, especially in the larger context of similarity avoidance as mandated by the OCP.

2.2.2.1. Gradient OCP Effects

Below are ranges of statistical significance. To put this in context, what it means to obtain each of these significance values is this: that the actual “0” value falls outside of (meaning either below or above) the central 99.9%, 99%, 95%, or 90% of the probability distribution of potential “0” values that the Monte Carlo procedure generates. Each of these ranges constitutes a confidence interval, and this is how the procedure is directly related to the determination of significance. Cases with extremely low 0 and E values are not meaningful to this analysis; that is because in those situations, where 0 has very small deviations from E, these deviations translate into very large differences in terms of the OlE ratio that results.

(9) Significance values: *** 00p< ** 001p< * l0 .05p< 7 0p. <0.10 (suggestive) n.s. .no notable effect

29 There is a gradient OCP[place] effect over the Gitksan lexicon. sequences which are homorganic (but not identical; C, C) are significantly underrepresented. This is seen in (10) (where cases of total identity are being left out, and significance is only noted for underrepresented pairs for visual ease).

As can be seen from above, Gitksan exhibits a gradient OCP[place] effect. While the magnitude of this effect is much smaller than is found in Arabic (compare with the table in (1) above), it is more along the lines of what is found in English (Berkley 2000). From the table in (10) we can see that homorganic labial pairs are extremely underrepresented, followed by dorsal pairs, and coronals. It is important to establish that these restrictions aren’t also found for heterorganic pairs of consonants. As the table in (10) indicates, consonants that differ in place of articulation are relatively unrestricted in terms of cooccurrence, as the OlE values are all around or above 1, with the exceptional cases discussed below. The statistically inconclusive value for glottal-glottal pairs is due to the fact that glottal consonants are extremely rare in roots, with only 3 observed occurrences of non identical glottal-glottal pairs. What is important here is how rare the glottal consonants I?, hi are in C position (in general) and/or in C position (in general) within the set of all

30 sequences in the lexicon. For example, 1Wis prohibited from appearing in C position (i.e. it is never found postvocalically; cf. discussion in Rigsby 1986), so only /hV?/ sequences would be counted (1?V’?Ipairs would not be counted since they constitute a case of total- identity, but this point is moot since no such forms are observed in the data). There can at most be 38 of these sequences, simply because I’?!occurs in C,, position in 38 of the 693 sequences in the database(and 1Woccurs in more than 38 times in C, position). This state of affairs, however, is extremely unlikely. The average under random chance is approximately 2.5 (i.e. the E value), and it is impossible for a value below this to be significant (and significance is what is needed to show an OCP[glottalj effect). Even if there were no observed occurrences of IhV?I, which would result in an O/E value of 0.00, this would still be far from significant. Since the numbers for the glottal consonants are so low in the database, perhaps more interesting is the behavior of glottalized consonants, which is discussed in section 2.2.2.2 below. Finally, it should be noted that dorsal-labial pairs and dorsal-glottal pairs (both in either order) are also underrepresented.’ While these are pairs of heterorganic consonants, the reason for the dorsal-labial6 underrepresentation is possibly due to the dual nature of segments like 1w! and of secondary articulations like labialization, which are specified as both labial and dorsal. To test whether these shared characteristics influenced these patterns of cooccurrence, the set of non-[w] labials and the set of non-labialized dorsals were tested for cooccurrence. Once these confounding factors are removed, labial-dorsal pairs showed non-significant deviation from 0 (61/70.2; O/E = 0.87, n.s.). The degree of underrepresentation is nearly the same as that found in (10) (O/E = 0.82), but no longer significant. What has been accomplished here is a reduction in the number of pairs under

16 It is worth pointing out that there is a potential danger in analyzing multiple individual statistical tests as is done in this chapter with the Monte Carlo procedure. Essentially what this procedure is doing is observing the o value for each individual class of consonant pairs and locating its value with regard to the 95% confidence interval (i.e. p < 0.05). However, falling outside of this interval means that a consonant pair would be expected to occur by random chance once very 20 samples; since there are 10 separate tests that are being run in (10), it is likely that one or two values may slip into the significant range by random chance. Thus, the labial-dorsal results could be a result of this “noise” in the data. The hypothesis here, however, is about what is involved in the diagonal cells of (10), which are predicted to have OlE < 1.00. Glottal-glottal pairs can be ignored, leaving three ‘tests’. If the Bonferroni correction (cf. Abdi 2007), a statistical procedure which makes adjustments in terms of significance for multiple hypotheses, is applied, then the criterial significance threshold becomes not 0.05, but 0.05/3 0.0167. As can be seen from the diagonal cells in (10), all three OlE values are below this threshold, validating the hypothesis. The same holds for (11), where the plosive-plosive cell would still be significant, but the other two diagonal cells would not.

31 consideration to the point where the same degree of underrepresentation is no longer significant. Taking the set {w, w’, k’’, kw, x”} yields a non-significant deviation from 0 (OlE = 0.57, n.s.). Thus, it still remains to be seen whether the labial glides and the labialized velars are “dual-category” in nature. The other potentially interesting result evident in (10) is the underrepresentation of glottal-dorsal pairs (OlE = 0.66, p < 0.05). It is possible that if the dorsal class is broken down into uvulars and velars, and it happens that the uvulars are actually the dorsal consonants that are cooccurring less frequently with the glottals, then there is a natural class being observed (the “guttural class”). This issue will be investigated further in Chapter 5. While on the topic of potentially ambiguous segments, it should be noted that this method of using cooccurrence restrictions in order to determine the class-hood of segment types is a unique approach to the investigation of phonological patterning. For instance, this is the approach used in some earlier works on feature geometry (McCarthy 1994), and several arguments similar to those here are made in (Frisch et al. 2004) and Coetzee & Pater (2006, 2008) for things like labio-dorsals and prenasalized stops based on cooccurrence facts. This method has now come to support what many approaches within generative phonology have claimed about how segments pattern categorically with respect to rules or constraints, and how distinctive features are employed in order to capture these patterns as natural classes. In sum, the prospect is that not only are natural classes defined by means of categorical behavior (such as inclusion as the structural description or environment of a rule, or target of a constraint), but also in terms of gradient behavior (as in cooccurrence restrictions). As noted above, the OCP effect in Arabic becomes evident once the class of coronals is subdivided into coronal obstruents and coronal sonorants (McCarthy 1988, Padgett 1995, Frisch et al. 2004). In order to test whether this same effect is present in Gitksan, the coronals were subdivided into obstruents and sonorants, and the obstruents were further subdivided into plosives (including both stops and affricates) and fricatives. These results are presented in (11).

32 ______

(11) Manner effects in coronal-coronal C,VC,,,sequences Obstruent Sonorant Plosive Fricative Obstruent Plosive 0 28** (3/11) t’u:ts’ ‘charcoal’ Fricative 1.11 — 0.39? (30/27) (2/5) ca+ ‘to lose us ‘finish’ (intrans)’ Sonorant 0.88 (n.s.) 0.84? 0.j? (44/50) (33/39.5) (19/25.5) nu:t ‘dress up’ lis ‘hang’ jal ‘to lie’

We can see from this table that within the coronal obstruents, plosive-plosive pairs are significantly underrepresented, while all other pairs failed to reach significance. It should be noted, though, that the patterning of the fricative-fricative pairs, as well as the sonorant sonorant pairs, is suggestive, since the numbers, while inconclusive, are in the direction predicted by an OCP effect. The E value for the CorFric-CorFric pairs is so low (which in turn may be why their O/E ratio is not significant) for a reason. Since there are only two coronal fricatives in the inventory, /s/ and IV, and since identical pairs are being excluded from consideration here, this means that only combinations of /s/ with IV are being observed. Effectively what this does is reduce the combinatorial possibilities for a particular individual C-C pairing (i.e. either s-+ or u-s), as opposed to an entire class of C-C pairings (such as CorPlos-CorPlos, where there are 5 distinct segments, yielding 10 different possible pairings). While most other individual pairings yield very low E values, the CorFric-CorFric pairing yields an E value of 5. The reason why it is this high is because of the overall high individual frequency of both /s/ and IV.

Among the dorsals, both plosive-fricative and plosive-plosive pairs are underrepresented, though the numbers for the plosives are much lower (see (12) below). The important point is the difference between the OlE values for same-manner pairs (plosive

33 ______

plosive) and the OlE values for the different-manner pairs (plosive-fricative): the former are more strongly underrepresented than the latter. The fact that both of these types are so strongly underrepresented is an indirect consequence of two factors: the “dorsal harmony” effect, to be discussed below in section 2.2.2.2., and the exclusion of the total identity pairs: 2/3 of the eligible types of fricative pairings, and 8/15 of the eligible types of plosive pairings, are velar+uvular pairs. These pairs are independently penalized by a pressure to have dorsals agree in place of articulation.

(12) Manner effects in dorsal-dorsal C,VC sequences Plosive Plosive 0.40** (8/20) gi:k’’ ‘to buy’ Fricative 0.66* inconclusive (13/20) (0/1.5) sqe: ‘be dark’ -

In order to determine whether there is a manner effect irrespective of place of articulation, all heterorganic pairs were counted and cross-categorized by manner and numbers were run for plosives, fricatives, and sonorants. In practice, if all of the heterorganic pairs of consonants are calculated together, and then the 2C position is reshuffled, homorganic pairs will be generated by accident every time the procedure is iterated. This is an undesirable result, as it allows place of articulation to methodologically sneak back into the mix and potentially bias the results. So in practice, the procedure had to be done piecemeal. First, the heterorganic CVC pairs were separated into groups: Cor NonCor, NonCor-Cor, Lab-{Dor, Glot}, {Dor, Glot}-Lab, Dor-Glot, and Glot-Dor. Each of these groups was then reshuffled individually, and the relevant 0 values were summed across all the groups for each iteration.

34 ______

(13) Non-homorganic CVC, sequences Obstruent Sonorant Plosive Fricalive Obstruent Plosive 1.02 (120/117.5) baq ‘to try, feel’ Fricative 1.02 i.72 (its.) (120/117) (10/14) lEak‘to bend hix ‘fat’ (trans)’ Sonorant 0.98 (n.s.) 1.01 1.07 (189/193) (51/50.5) (40/37.5) mit ‘full’ m’as ‘grow up limx ‘song, (p1)’ sing’

As can be seen in (13), there were no significant effects found for any of the non-homorganic combinations of manner. The behavior of these classes also fits the pattern found in Arabic, as well. In Arabic, the manner cooccurrence restrictions (illustrated in ex. 1) are modulated by place of articulation; no effect is found for non-homorganic combinations of consonants with the same or similar manner (Padgett 1995, Frisch et al. 2004). From this we can conclude that there is a slight influence of manner on the overall similarity of consonants, however, the manner effect is only present within each place distinction; overall, there is no manner effect without a place effect. In other words, the fact that a consonant pair matches in terms of manner contributes to the similarity between the two consonants, and the OCP[place] effect is modulated by similarity, but it is fundamentally an OCP[Place] effect, as there is no evidence to suggest it is an OCP[manner] effect per se. This has implications for This is an extremely important point. The fact that manner features have no independent effect, but rather contribute only as subsidiary features to place features indicates that this is a genuinely phonological phenomenon’ i.e. a grammatical effect. For instance, if the patterns observed above were7simply a phonetic effect, then it would be expected that place of articulation and manner, of articulation (or potentially any feature) 17 Or in theveryleast the effect is predicated over articulatory gestures, and that manner gestures are simply detailed aspects of gestures that are executed by the same articulator; hence, there would be no “detailed” effect without the gross effect.

35 would contribute equally to the effect; this however, is not the case. The effect seems to be modulated by the different kinds of constraints that the grammar provides; i.e. constraints on place cooccurrence, and constraints on place cooccurrence within manner classes, but no atomic constraints on manner cooccurrence (the exact constraints to be employed are discussed and defined in Chapter 3). In other words, these effects are to some extent modulated by Universal Grammar. So far we have been concerned only with local consonant pairs; that is, only pairs of consonants separated by a vowel (CXVCY).However, when the context is extended from transvocalic (CXVCY)to long-distance (Cx...C.. .C) pairs, the OCP effect found above completely evaporates. This is illustrated in (14), where we can see that there are no OlE values significantly different from 1.00 for any homorganic pairs of consonants.

(14) No OCP[place] effect in long-distance C. . . C.. .C sequences

Labial Coronal Dorsal I Glottal Labial inconclusive (2/3) k’ibilap ‘gravel, small rocks’ Coronal 1.00 1.08 (53/53) (109/100.5) maqt ‘to leave sg”at ‘tojoke’ (trans)’ Dorsal 0.96 (n.s.) 0.92? (29/30) (130/142) mumq’ ‘to ci1ks ‘to melt smile’ (intrans)’ Glottal 1.30 0.90 (n.s.) 1.13 inconclusive (6/4.5) (28/31) (20/18) (1/1) hawaw hilin ‘be lonely’ hanix ‘to be haloLlo ‘cloth, ‘mountain lion’ thin’ sail’

It can be concluded, then, that the OCP effect found in Gitksan roots, whereby there is a general avoidance of homorganic consonants in transvocalic contexts, is not found in any long-distance contexts. Thus, proximity plays a role in the place cooccurrence effect.

36 Gitksan can be compared with Arabic in this respect. In Arabic, the same asymmetry is present, with an extremely strong local effect and a weak long-distance effect (Frisch et al. 2004). The same result obtains in English, where the OCP effect diminishes with increased material intervening (Berkley 2000). In Gitksan, the overall magnitude of each effect is dampened: there is a weak local effect, and no reliable sign of a long-distance effect. It appears that the difference between Arabic and English on the one hand, and Gitksan on the other, is simply one of degree. Assuming that cooccurrence constraints come in different locality thresholds, and that the more “local” versions of a constraint are in a subset relation with respect to less “local” ones (Suzuki 1998), then a prediction to be made is that all of the individual OCP[place] effects that are seen above for transvocalic CVC contexts should be found in adjacency contexts, as well. If it is also assumed that there is a universal markedness hierarchy that determines the relative ranking of OCP[PLAcE]constraints (such that OCP[lab] >> OCP[dor] >> OCP[cor]), then the effects should overall be greater in magnitude in adjacency contexts, but the strength relations among the different types of OCP[place] constraints should be roughly equivalent in these adjacency contexts. Values for adjacent homorganic pairs are displayed in (15). Since it was shown that identical pairs are subject to potentially independent constraints in adjacency contexts (i.e. constraints banning geminates), these pairs have been removed from the counts.

(15) Adjacent homorganic pairs, by place (identicals not included) labial-labial coronal-coronal dorsal-dorsal 0.00 n.s. 0.82** 0.08*** (0/0.92) (57/69.9) (2/24.6) - k’”ast ‘be broken’ q’esxq ‘be green, unripe’

Labial pairs exhibit non-significant results (with 0 as the 0 value), though the findings in transvocalic contexts are fully consistent with these numbers. The coronals and dorsals are underrepresented. While the coronals are nearly identical to the transvocalic cases, the dorsals are greatly reduced (OlE = 0.54 for transvocalic contexts), as would be expected. This illustrates how the relative markedness of different places of articulation in

CXVCYcontexts is preserved in adjacency contexts.

37 Since “the coronal-coronal effect” is nearly as weak as in the transvocalic cases, coronal place was broken down by manner:

(16) Adjacent coronals by manner CorObs-CorObs 0.82** 0.49 (92/113.2) (1/2) gwa:st ‘to lend’ &ajn ‘Chinese’

While the coronal sonorants are underrepresented roughly as expected (O/E = 0.60), the coronal obstruents remain quite high (O/E = 0.81). Therefore, the class of coronal obstruents was again further broken down into plosive-plosive and fricative-fricative pairs:

(17) Adjacent coronal obstruents by manner

CorPlos-Cor-Plos - CorFric-CorFric 0.00 0.00*** (0/0.7) (0/6.7)

Broken down into these categories, it can be seen that the restrictions against adjacent coronal plosive pairs and adjacent coronal fricative pairs are absolute: there are no exceptions. Since pairs of coronal obstruents yield an O/E value of 0.82, this illustrates the point that there is a gradient restriction on pairs of adjacent coronal obstruents, but an even tighter (absolute) restriction on pairs of adjacent coronal obstruents that agree in continuancy. The generalization, then, is that adjacent coronal obstruents are tolerated (as there are 92 such pairs in the database, and the O/E value is the same for all adjacent coronals), but adjacent coronal obstruents agreeing in continuancy are not tolerated. Interestingly, the prohibition on pairs of coronal consonants remains rougly the same in CXVC,,contexts as in adjacent contexts: O/E 0.83 for the former, 0.82 for the latter (and 0.82 for adjacent coronal obstruents, as well). While the discussion so far has been focused on dissimilatory effects, we turn next to a somewhat surprising result: gradient assimilatory effects.

18 Of course, this value would be 0.00 (just as in 17) if[j] and [j’]were not classified as coronals.

38 ______

2.2.2.2. Gradient Assimilatory Effects

While the OCP-type of dissimilatory effects found above are not unexpected, there is also an assimilatory effect found in the data. It was noted above that dorsal-dorsal pairs are significantly underrepresented in the Gitksan lexicon (OlE 0.54 in transvocalic contexts). However, when we look within the class of 9dorsals’ we find an altogether different tendency. By opening up the class of dorsals, what, effectively happens is an “undoing” of the overall OCP [dors] effect; that is, opening the class abstracts away from that overall effect. What “opening up the class” of dorsals means, exactly, is this: that cooccurrences among the dorsal consonants are being counted within the set of all dorsal-dorsal CVC pairs. What this means is calculating new E values relative to this subset of CVC pairs. This being the case, the E value for an (ordered) sequence like IkVx/ is calculated by taking the

frequency of 1k!as Ci of a dorsal-dorsal CVC pair, multiplied by the frequency of !xI as the 2C of a dorsal-dorsal CVC pair, multiplied by the total number of dorsal-dorsal CVC pairs, and so forth. Among the velars and uvulars, which make up the dorsals, we find that velar-uvular pairs are underrepresented (OlE = 0.57), and that uvular-uvular pairs are roughly as expected by chance (OlE = 1.20):

(18) Dorsal agreement in dorsal-dorsal sequences Velar Uvular I Velar 088(ns)

guxs ‘wake up’ Uvular 0.57* 1.20 (n.s.) (10/17.5) (6/5) hoga ‘be like’ Ge:x ‘to slide’

19 Looking within the class of coronals would more than likely also prove interesting. While there are no subsidiary place contrasts within the coronal class in the inventory in (2), there are dialects of Gitksan which ethibit an s/$ distinction to some degree (see discussion in Rigsby 1986). If there happened to be a dialect which fully contrasted the two, then values for anteriority could be investigated. Even if this were the case, though, the numbers of these segments in the lexicon would probably not be very high. An analogous case would be to look at the sibilant obstruent vs. lateral obstruent distinction, as Hansson (2001) notes diachronic cases of long-distance assimilation between these segment types. Future dialectological research will prove interesting for both of these areas.

39 What this indicates is that although dorsal-dorsal pairs are underrepresented as a class, when dorsals do occur together, there is a preference toward similarity. This appears to be something of a gradient consonant harmony (Hansson 2001, Rose & Walker 2004). Interestingly, this assimilatory effect remains when the dorsals are examined in long-distance

(Cx...C. . .C) sequences:

(19) Dorsal agreement in dorsal-dorsal long-distance Cx.. .C. . .C, sequences Velar - Velar 1.39 (n.s.) (8/6) x”dak’’ ‘to shoot (intrans)’ Uvular 0.66* l.8i-” (15/23) (18/10) xsoo:q ‘be first’ sinq ‘to disbelieve’

In fact, the effect becomes even stronger in terms of the overrepresentation of uvular-uvular pairs (O/E = 1.83). What this indicates is that this gradient consonant harmony (or “uvularity harmony”) extends beyond transvocalic environments, and can be viewed as a type of gradient long-distance consonant agreement (Hansson 2001, Rose & Walker 2004). Since intervening vowels are lowered adjacent to uvulars (as noted in Chapter 5), it is likely that these vowels have an effect on the overall pattern. For instance, uvularity (or more precisely its F 1 correlate) could have been (historically) transmitted from i to 2C (or vice versa) by intervening vowels. Since uvular consonants lower adjacent vowels (a phenomenon which is discussed at length in Chatper 5), it is possible that the raised F 1 of a vowel lowered by a uvular could be reinterpreted by a listener/learner on a more distant dorsal, with the result that the dorsal is interpreted as uvular. Similar cases are discussed by Hansson (2007a) with regard to secondary articulation agreement, and the general idea that sound changes such as this result in synchronic sound patterns can be found in works such as Blevins (2004), Hansson (2001), etc. The role of place and vowel quality in long-distance contexts, especially with regard to dorsal harmony, will be discussed further in Chapter 5. A methodological point must be made here: if all of the Expected values in (18) are summed (7+5+17.5=29.5), and all of the Observed values are summed (6+6+10=22), they result in differing values. Given that these values are supposed to cover all Dorsal-Dorsal

40 pairs (and no other pairs are included), then what we end up with is a discrepancy (22 vs. 29.5). The discrepancy is due to the fact that all of the identical pairs have been removed from these counts, meaning that after each re-shuffling of the Monte Carlo procedure all identical pairs that arose from the reshuffling procedure were thrown out (though the particular i and 2C instances were kept available for the next shuffle). This being the case, the identical dorsal-dorsal pairs are extremely overrepresented (among dorsal-dorsal pairs), as O/E = 1.88 (13/7). In other words, many more non-identical pairs (shown here) are expected relative to identical pairs (not shown here), as compared to the number of observed pairs, where there are not many non-identical pairs (shown here) and unexpectedly many identical pairs (not shown here). This indicates that the gradient uvularity harmony described above is actually stronger than it originally appeared, if cases of identical consonants are inserted back into the total counts. While cases of complete identity often serve as exceptions to dissimilatory restrictions (and thus removed from the data examined), what is being observed here is a case of assimilatory behavior. That is, consonants that are more similar are found more frequently than those that are less similar. Interestingly, the OlE value for identical dorsals is even higher than that of identical labials (OlE = 1.16). This difference in phenomena stands as justification for putting the cases of identical pairs of consonants back into the total counts. There is also a “philosophical” argument for including the total-identity pairs. In principle, a pair of consonants that are identical can be an exception to an OCP[F] constraint, since the pair will agree for the feature [F], and by definition violate the constraint. However, a total- identity pair cannot by definition be an exception to a constraint that demands agreement with respect to a feature [F]. This is because the identical pair will satisfr the agreement constraint. Thus, since this appears to be an agreement constraint at work (to be outlined in chapter 5), and since total-identity pairs cannot by definition behave exceptionally with respect to this constraint, it is not reasonable to leave these pairs out of the counts. After putting the identical pairs back in, it can be seen that the same basic trend holds; (20) illustrates what the OlE values look like once the identical dorsal pairs are added back into the counts (specifically the 0 values), both for the actual lexicon (which equals 0), and for the 10,000 Monte Carlo lexica (whose 0 values cluster around the E value).

41 (20) Dorsal agreement in dorsal-dorsal CVC, sequences (identical pairs back in) Velar - - Uvular

Velar 1 40* (13/9.3) •iIiiirI*i guks ‘to wake up’ Uvular 0.57* 1.46* (10/17.5) (12/8.2) Ga:k”’ ‘sinew’ xo:x‘to yawn’

The same holds true for long-distance dorsal-dorsal consonant pairs once the identical pairs are pooled back in:

(21) Dorsal agreement in dorsal-dorsal long-distance C,.. . C. . . C’,,sequences (identical pairs back in) Velar Uvular 49* Velar 1 (18/12) gamk’tobehot’ 1 Uvular 0.66* 1.24* (15/23) (20/16.1) gloq ‘shame’ Gane:xs ‘bridge’

To reiterate what was said earlier, what is being observed here amounts to a type of “uvularity” harmony whereby dorsal pairs tend to agree with respect to uvularity. That is, velars tend to cooccur with velars, and uvulars tend to cooccur with uvulars, leaving “disharmonic” velar-uvular pairs significantly underrepresented (O/E 0.57 transvocalic, 0.66 long distance). In addition to the assimilatory behavior of the dorsals, there is an additional assimilatory tendency found among glottalized consonants. While nonglottalized nonglottalized consonant pairs are relatively what we would expect them to be (O/E = 0.94), the glottalized-nonglottalized pairs are slightly, but significantly underrepresented (O/E = 0.84). Furthermore, glottalized-glottalized pairs are significantly overrepresented (O/E = 1.47):

42 ______

(22) Laryngeal- - agreement in all sequences (obstruents and sonorants alike) - Plain Glottalized Plain 0942 (106/113) ban ‘to ache’ Glottalized 0.84** 1.47* (105/125) (20/13.5) gwina ‘ask for, beg’ t’e:l’t ‘be fast’

This only presents part of the picture, though, since identical pairs of consonants were originally removed from these counts. As with the dorsals, what we have here is a case of assimilatory behavior. That is, consonants that are more similar are found more frequently than those that are less similar. Just as in the case of the dorsals, this difference in phenomena stands as justification for putting the cases of identical pairs of consonants back into the total counts. After doing so, it can be seen that the same basic trend holds:

(23) Laryngeal agreement in all C,VC sequences (obstruents and sonorants; identicals in) Plain Glottalized Plain (419/409.6) wit_collarbone Glottalized 0.89** 1.50* (152/169.8) (23/15.5) loots’ ‘elderberry’ t’aw’ilt ‘fishhook’

Again, once the identical pairs are placed back into the data, we see that

nonglottalized-nonglottalized pairs have an O/E value around 1, and that glottalized-plain pairs are slightly, but significantly underrepresented. It should be noted that the overrepresentation for the plain-plain pairs is extremely slight; the reason that the overrepresentation is significant is due to the sample size, where the number and frequency of non-glottalized consonants is relatively much larger. Finally, glottalized-glottalized pairs are very overrepresented (though the difference is only 7.5 items), with an O/E value of 1.50. These results indicate that once the identical pairs are counted back in, the effect becomes stronger.

43 These results show that there is a significant overrepresentation for the “agreeing” pairs, in particular for the glottalized consonants, and a significant underrepresentation for the “disagreeing” pairs. For the latter, it is important to note that the effect itself isn’t very large (only 11% fewer disagreeing pairs than would be expected by chance, as OlE = 0.89), though the effect is significant, and to put it into perspective, it is roughly at the same level as the overall OCP[cor] effect. Going back to a point made in Section 2.2.1, the plain/glottalized consonant patterns here nicely illustrate the difference between raw frequency (0 values) and OlE ratios. There are 152 occurrences of CVC’ or C’VC pairs, but there are only 23 total occurrences of C’VC’ pairs, meaning that the CVC’IC’VC pairs are higher in raw frequency than C’VC’ pairs. Nevertheless, the former are underrepresented (suggesting that they are penalized in the phonology), while the latter are overrepresented (suggesting that they are favoured in the phonology). If raw counts alone were the only counts being observed, then the opposite interpretation would be drawn. Earlier, for OCP effects, the totals were relativized to a particular subcategory of consonants. The same procedure was done for the case of dorsal agreement, where the subclass was the “dorsal” place. Because of these previous findings, it is relevant to do the same thing here for glottalization agreement, in order to test whether there is another subcategorization that is responsible for the effect. The primary cut is along major class lines: obstruent vs. sonorant. Then, within the obstruents, the natural cut is along plosive vs. fricative lines. This cut is especially important, as all fricatives are predictably non glottalized. In order to see whether major class features had any effect, consonant pairs were first broken down by major class, and then broken down by glottalization. (24) gives the results for pulmonic vs. ejectives:

44 (24) Laryngeal agreement in all CVC sequences (obstruent pairs) Pulmonic Cilottalized

Pulmonic 1.04** (161/154.3) sa:k ‘oolachen, candlefish’

Glottalized 0.84** I.8o (72/85.3) (15/8.3) q’esi ‘knee’ t1’o:k’ ‘mud’

It can be seen for the obstruents that the overrepresentation for pairs agreeing in laryngeal features is significant for both pulmonic-pulmonic (OlE = 1.04) and glottalized-glottalized pairs (OlE = 1.80), and significant for both. The mixed pulmonic-glottalized pairs are significantly underrepresented (OlE = 0.84). In order to see if maimer had any effect, the consonants were broken down further into plosives (by removing the fricatives, which are not glottalized), and then broken down into pulmonic vs. glottalized; this is illustrated in (25).

Since the fricatives were removed, we see that the effect remains comparable, with agreeing pulmonic-pulmonic (OlE = 1.13) and glottalized-glottalized (OlE 1.75) still significantly overrepresented, and disagreeing pulmonic-glottalized pairs (OlE 0.75) still significantly underrepresented. If the OlE values are tracked for disagreeing pairs from C-C to Obs-Obs to Pbs-Pbs pairs, we can see that they decrease. In other words, the underrepresentation increases in magnitude. This is exactly the type of trend that would be expected if the effect were based on similarity (cf. Frisch et al. 2004). Since the effect was consistent for the obstruents, the sonorants were also investigated. This is found in (25):

45 (26) Laryngeal agreement in all sequences (sonorant pairs) Plain Glottalized Plain 0 99 (n s) (50/50.6) wan ‘deer’ Glottalized 1.06 (n.s.) 0.64 (n.s.) (21/19.9) (1/1.6) j’ans ‘leaf, grass, weeds’ n’uw’ ‘to die’

As it happens, there was no significant effect found for any of the sonorants (though the numbers for glottalized-glottalized pairs are very small). This indicates that the gradient agreement effect for glottalized consonants is limited to only the plosives, and that glottalization on sonorant consonants may be fundamentally different from glottalization on obstruents (i.e. ejectives). There are at least two things that would motivate this difference. The first is the fact that there is a difference in timing and articulation of the glottal gesture between glottalized sonorants and ejectives. Glottalized sonorants tend to have the glottal gesture timed during modal voicing of the sonorant, while ejectives tend to have the glottal gesture timed after the release of the plosive. The second is the asymmetry in deglottalization found in reduplicative contexts across the Tsimshianic family, whereby in one language (Nisgha) glottalized sonorants in reduplicant onsets are deglottalized, but ejectives in the same position maintain their glottalization. This phenomenon is discussed at length in Chapter 6. Finally, since it was found above that the dorsal agreement effect held in the long- distance cases, this was also tested for the pulmonic/glottalized pairs. This is shown in (27), where all plosives were observed in Cx.. . C. . . C, sequences:

(27) Laryngeal agreement in all Cx.. . C. . . C, sequences (plosive pairs, long-distance) Pulmonic Glottalized Pulmonic 1.01 (n.s.) 9: jjf (46/43 7) Gidalq ‘crane’ Glottalized 0.96 (n.s.) 1.64 (n.s.) (15/15.7) (1/0.7) q’esxq ‘green, unripe’ k’ak’jot1’ ‘ankle’

46 We can see here that unlike the dorsal agreement, the glottalization agreement does not

persist in long-distance contexts — none of the OlE values for agreeing or mixed pairs is significantly different from 1.00. This, however, may be the result of skewing by a lopsided distribution of glottalized plosives in the long-distance plosive-plosive pairs. Out of these 62 total pairs, 59 of these contain a non-glottalized plosive as the second consonant, regardless of whether the first consonant is glottalized or not. This skewing can affect the statistical method, in that the re-shuffling of the second consonant of all 62 plosives in that position will only derive 4 possible results: all three glottalized s2C matched with a glottalized ,1C or two glottalized s2C with glottalized ,1C or one 2C with glottalized ,1C or none. Thus, all trials of the Monte Carlo procedure will result in four possibilities, with the actual lexical count having only one of the glottalized s2C paired with a glottalized .1C The underlying reason for the lopsided distribution of glottalized consonants may be due to asymmetries in the positions where these consonant types may occur (i.e. a directionality effect), or it may be due to phonotactics in general (i.e. properties of glottalized consonants in clusters) as opposed to the special type of “cooccurrence” phonotactics discussed in this dissertation. The general effect of gradient laryngeal assimilation in glottalization discussed in this section is somewhat surprising, as there are abundant cases of glottal dissimilation in the literature (MacEachem 1999), as well as well-known cases of synchronic and diachronic processes involving dissimilation (Fallon 2002) such as Grassmann’s Law, observed in Sanskrit and Ancient Greek, and which states that if two aspirated consonants are in adjacent syllables, the first loses its aspiration. There do exist cases of laryngeal agreement; however, these cases are more rare than laryngeal disagreement, and they are typically not gradient in 20nature . Overall, these special cases of gradient consonant agreement will be discussed in more detail in Chapters 4 and 5, where they will be given a formal analysis using the Agreement by Correspondence paradigm (Hansson 2001, Rose & Walker 2004).

20 Though see Hansson (2001), Rose & Walker (2004), and Rose & King (2007) for some cases of gradient consonant agreement. These cases will be dealtwith in more detail in Chapter 4. There isa good chance that a gradient laryngeal disagreement may commonly exist cross-linguistically, but go unnoticed and not be accurately recorded as such.

47 2.3. Summary and Discussion

We have seen that there is an OCP effect shaping the frequency distribution of homorganic consonant pairs in the Gitksan lexicon (excluding cases of identical consonants, which are exempt from the effect). We have also seen that there is a contribution of manner to the effect, as pairs of coronal plosives are underrepresented, while pairs of coronal sonorants, pairs of coronal fricatives, or any mixed pairs are not underrepresented. Within long- distance contexts, however, the OCP effect disappears. We have also seen that there is a surprising assimilatory effect that is embedded within the larger dissimilatory context: within the class of dorsals, there is a tendency for uvulars to cooccur with uvulars and velars to cooccur with velars, and this effect extends to long-distance cases. The same is true for glottalized-glottalized pairs and pulmonic-pulmonic pairs of plosives for transvocalic contexts. An immediate question to be raised is, are these patterns phonological? In other words, should they be analyzed in the phonological grammar alongside other categorical phenomena? The claim here is that the gradient phenomena presented in this chapter are phonological. There are four lines of argumentation to support this claim. The first is that the patterns outlined above are statistically significant. Given this, it stands that these sound patterns should be explained by some linguistic mechanism. Secondly, one of the most convincing reasons has to do with the place/manner dependencies mentioned above. In Gitksan (as is reported in many other languages displaying these gradient OCP effects), there is a place of articulation effect. There is also a manner effect, but only as modulated by the place effect. Manner features do not contribute an effect independent of place features. If everything else were equal, then any kind of features would be expected to contribute to an OCP effect; however, this is not the case. It is reasons like this that motivate researchers like Frisch et al. (2004:223) to claim that the overall effect is phonological in nature, and hence the lexical frequency distributions should be reflected somehow in the phonological grammar. Third, the evidence from the behavior of the labialized velars speaks strongly to the phonological nature of these consonant cooccurrence effects. Again, if this were just a phonetic effect, it would be expected that labialized velars would pattern equally with labials and dorsals (just as would the labio-velars); instead, these segments pattern with the dorsals, and not the labials, suggesting that primary place features are responsible for the way these

48 segments are categorized. Finally, the phonological status of these effects can be argued on the basis of their exceptional nature; or rather, the fact that many “categorical” phenomena in phonology do in fact have exceptions. One or two exceptions seem to be tolerated by phonologists, and it could be argued that even three or more exceptions would not place the “categorical” status of a phenomenon in jeopardy. But is there a clear line to be drawn between a categorical pattern with some exceptions and a gradient pattern? This line of argumentation leads back to the first point made, that the patterns mentioned above are statistically significant. That being said, significance will be the deciding criterion in the determination of phonological versus non-phonological patterns in this thesis. The results so far also lend themselves toward a perceptual/processing explanation. It is assumed by many that complete identity cases and heterorganic cases are not subject to an OCP effect, and only with increased similarity are consonant pairs subject to stronger effects. The idea is that perceiving or processing two things in close proximity that are completely identical is easy, and doing so for two things that are completely different is also easy. However, perceiving or processing two things that are similar is difficult (Pierrehumbert 1993, Frisch 1996, Hansson 2001, Frisch et al. 2004, Rose & Walker 2004). This perceptual/processing account jibes well with the patterns noted above for Gitksan in that a) there is an identity effect and b) the degree to which segments are similar decreases their likelihood of cooccurring. In order to determine if the processing/perceptual account is actually responsible for these effects (both in Gitksan and in other languages), further instrumental studies would need to be performed, which are outside the scope of this thesis. Thus, at the moment, we are left to speculate. In languages that exhibit similar OCP-type patterns, it has been shown that these patterns are a part of a speaker’s internalized knowledge (Frisch & Zawaydeh 2001, Rose & King 2007). In an ideal world, these types of tests could be performed on speakers of Gitksan; however, this is beyond the scope of this thesis. Rather, it will be assumed that this will be true of Gitksan, just as it is with other languages exhibiting similar patterns.

49 Chapter 3: Constraints on Gradient Lexical Phonotactics

3.1. Introduction

Capturing gradient phenomena like those found in Gitksan has been problematic for approaches like Optimality Theory (Berkley 2000, Frisch et al. 2004). This chapter attempts an account for these patterns in Harmonic Grammar, a close relative of OT, and relates the findings in the results section of Chapter 2 to the idea of constraint ranking and constraint weighting. This chapter also discusses the implications and predictions of models which employ weighted constraints in their architectures, especially with regard to the Gradual Learning Algorithm (Boersma 1998). The claim put forward is that statistical patterns (based on lexical frequencies) constitute phonological knowledge, and that this knowledge is encoded in the weighting of constraints.

3.2. Background on Theoretical Approaches

Given the gradient nature of the phonotactic patterns found in the Gitksan lexicon, when giving a formal analysis it becomes necessary to consider what it is that is being modeled when the phonological grammar is being appealed to. It is also necessary to consider what the grounds would be for rejecting some of the competing theoretical approaches, as well as what the grounds are for preferring a given approach. This section will lay out two general approaches, the lexical indexation model of Coetzee & Pater (2006), and an approach couched in Harmonic Grammar (Coetzee & Pater 2008). After presenting the crucial problem facing a theory with strict domination of constraints, these two approaches will be discussed. At this point, it is relevant to question whether there is even anything for a grammar to model in the data. In other words, should the patterns observed across the Gitksan lexicon be encoded in the phonological grammar? The claim here is that there are several good reasons for locating these effects in the phonology of the language, and not as random comfigurations of the lexicon. The arguments for the phonological status of these patterns are

50 presented in section 3 of chapter 2, but the most immediate reason would be that the patterns discussed above are statistically significant. This shows that each pattern is not merely the result of random chance acting on the distribution of consonants in the lexicon. Rather, these are robust patterns. Another reason for taking this point of view comes from the psycholinguistic literature on OCP effects. Several researchers have found that speakers have an internalized knowledge of the gradient patterns found over the lexicon (see, for example, Greenberg & Jenkins 1964, Treiman et al. 2000, Bailey & Hahn 2001, Frisch & Zawaydeh 2001, Coetzee 2004, 2008, Hammond 2004). This aspect will be discussed further in section 3.3.1. One question is, if these patterns are to be represented in the phonological grammar, what is the exact mechanism that is responsible for them, or alternatively, what guides the grammar to take the shape that it has (i.e. reflecting the patterns in the lexicon)? One way of accounting for gradient phonotactics is by an independent metric of similarity, as has been advocated by Frisch et al. (2004). Under this approach, the more similar two consonants are, the less likely they will cooccur. Formally, this means that there is a correlation between similarity values and OlE values. Frisch et al. (2004) develop a similarity metric whereby shared natural classes determines the degree to which two phonemes are similar. The metric is formalized as such:

(1) Natural classes similarity metric (Frisch et al. 2004) natural classes containing both x andy Sim(x,y) = natural classes containing either x or y (or both)

Frisch et al. use this metric to determine how similar two consonants are to each other. Their claim is that the higher the similarity, the less attested a given pair of consonants will be, and they make this point with data from Arabic. While this metric provides a tight fit to the Arabic data, Brown & Hansson (2008) have shown that in Gitksan the correlation of OlE with numerical similarity does not provide as tight a fit for the labials and coronals as would be expected. Coetzee & Pater (2008) make the same point for Muna, but with respect to the dorsals vs. coronals. Furthermore, Brown & Hansson note that if labial-labial and coronal-coronal pairs are compared, they yield

51 extremely close similarity values, but drastically different OlE values. This is illustrated in (2), where these results are compared to Arabic.

(2) Average similarity values for Gitksan and Arabic (Brown & Hansson 2006) labial-labial T coronal-coronal Gitksan 0.26 0.24 Arabic 0.32 0.21

While the similarity values for labial-labial and coronal-coronal pairs are nearly identical, the O/E values for each are 0.16 and 0.83, respectively. The similarity-based approach would not predict such a great difference in OlE value; rather, it would predict O/E values for labial-labial and coronal-coronal pairs that are approximately the same. In contrast to this similarity-based approach are constraint-based approaches. In the following sections, the problems that gradient phonotactis poses for Optimality Theory are laid out, and two constraint-based approaches to these problems are presented. These include a mechanism of lexically indexing constraints to lexical strata, and employing weighted (instead of strictly dominated) constraints.

3.2.1. Problems for Optimality Theory

It has been pointed out by Frisch (1996), Berkley (2000), and Frisch et al. (2004), among others, that there is an inherent difficulty in accounting for these types of gradient phonological patterns within a framework like standard Optimality Theory (Prince & Smolensky 1993 [20041). The problem lies in the inability to capture a gradient pattern with a set of strictly ranked constraints. Any given ranking of a set of constraints will result in categorical output patterns. This is illustrated in the case (of a hypothetical word) below, where a constraint prohibiting cooccurring coronals (OCP-coR) is in conflict with a faithfulness constraint mandating identity between input and output (FAITH):

(3) Enforcing OCP-coR /sitJ OCP-COR FAITH sit crsjp *

52 In this case, faithfulness is violated in favor of respecting OCP-coR. Under this configuration, it would not matter whether the underlying form was /sit/ or /sip/; the highly ranked markedness constraint OCP-coR dictates what the surface form will look like:

(4) Enforcing OCP-coR (assuming a rich base) /sip/ OCP-cor FAITH sit c sip

If the ranking was reversed, as in (5), then the OCP would never have any noticeable effect, since underlying forms would always surface unaltered, whether they violate OCP-coR or not:

(5) Blocking OCP-coR /sit/ FAITH OCP-COR crsit * sip *!

While this particular ranking allows for forms like [sit] to surface, it does not allow for an active OCP effect to be felt, counter to what has been demonstrated for Gitksan. This simple example illustrates that gradient phonemena like those found in Gitksan are a problem for a theory with only strict domination of constraints. However, if faithfulness is highly ranking and all lexical forms are identical to their output forms (by lexicon optimization, Prince & Smolensky 1993[2004]), then the OCP effect can be “preserved” in each lexical item (i.e. all forms surface consistently without variation, and the sum of all lexical items exhibit the gradient OCP pattern), though expressed nowhere in the grammar. This is again a problem for OT and strict domination. The problem is further exacerbated when the leamability of the system is brought into the picture. Under standard models of learning in OT, such as the Constraint Demotion algorithm (Tesar and Smolensky 2000), or Biased Constraint Demotion (Prince & Tesar 2004), the grammar has no way of tracking the OCP violations of fully faithful surface forms. Rather, during the course of learning, those markedness constraints that the learner encounters violating structures in the learning data are demoted in ranking until convergence is reached (convergence being the state where no further learning takes place because the grammar is consistent with all output forms in the

53 language), and learning ceases without regard to the statistical pattern that a lexicon may exhibit. Under these algorithms and in a case like Gitksan, faithfulness constraints will come to dominate all relevant markedness constraints.

3.2.2. Lexical Indexation

One possible approach to gradient phonotactics is to attempt to “derive” the patterns in the lexicon from the grammar, or in other words, to have the grammar replicate the distributional patterns that are observed. On this view, the assumption is that if all types of C-C combinations are exactly equally frequent as inputs in a “rich” base (i.e. not the set of inputs that make up the actual lexicon of the language in question), and this distribution is submitted to the grammar, then the result would be a skewed distribution (i.e. the gradient phonotactic patterns that are observable on the surface). This distribution would include some combinations of consonant pairs that would be statistically underrepresented. This is essentially the approach taken by Coetzee & Pater (2006), and the way that such an approach would work is outlined below. The challenge is to capture gradient patterns with only the mechanism of constraint 21ranking One possible approach is to employ probabilistic rankings of constraints .(Boersma 1998, Hayes & MacEachern 1998, Boersma & Hayes 2001, Hammond 2004). Under this approach, rather than being fixed in their ranking, constraints are assigned a ranking value located on a continuous scale. At the time of evaluation, noise is added to the ranking value (where “noise” refers to a probability distribution along the scale). The result is the possibility for one constraint to outrank another constraint at evaluation a given number of times, but to have the opposite ranking some percentage of the time. This stochastic approach can correctly derive the frequency of OCP application vs. non-application (for example, very high frequency for labials, moderate frequency for dorsals, and lower frequency for coronals), but as Coetzee & Pater (2006) point out, this approach makes the prediction that any given lexical item will undergo the effect a certain percentage of the times it surfaces, and that it will also not undergo the effect a certain percentage of times. For the phenomena at hand, this prediction is incorrect. For example, the lexical item /pitl ‘divorce’

21 There could of course be other mechanisms available, such as those involving leamability. One such mechanism will be discussed in Chapter 4.

54 will consistently surface with consonants that disagree in place, while /k’e:q/ ‘to run away, flee’ will consistently surface with dorsal consonants; there is no repair strategy that will be employed in order to avoid the OCP violation at any time. In other words, the lexicon systematically contains more cases like /pit/ ‘divorce’ than expected, and fewer cases like /k’e:q/ ‘drill’ than expected. In order to account for this type of lexical gradience, Coetzee &

Pater (2006) employ a set of lexically-indexed faithfulness constraints (cf. Pater 2000, Ito & Mester 2003). Lexical items are indexed to these faithfulness constraints, which protect them from application of the OCP. Following Coetzee & Pater, these constraints will be termed Fl, F2, F3, etc. In this type of context, the role of the phonological grammar is diminished, as lexical indexation essentially prevents the grammar from having any categorical effect. It is perhaps more instructive to first present (hypothetical) cases where the OCP is respected and dissimilation obtains. In this way, the role of the principle of Richness of the Base (Prince & Smolensky 1993 [20041) can be highlighted. In these cases, lexicon optimization would force the learner to posit the output form as the input, rendering any type of OCP effect moot. (It is important to note here that since this is an exercise in modeling a grammar with an assumed rich base, the hypothetical examples used in the tableaux from here out should not be taken to be real lexical items. Likewise, since we are only concerned with whether the OCP applies or does not apply, any type of repair which an output form undergoes as of the result of respecting the OCP is not of interest here). The tableaux in (1) and (2) illustrate the enforcement of an OCP constraint, and (3) illustrates the reverse ranking and blocking of the OCP effect. Thus, in order to allow for forms like [sit], which violate OCP-coR, to occasionally surface, an additional faithfulness constraint must be introduced: Fl. This constraint can be indexed to specific lexical items, and when ranked above OCP COR, will protect them from the OCP effect by mandating identity between input and output forms. This is illustrated in (6).

55 (6) The role of F1 in blocking the OCP effect a. /sit/i Fl OCP-coR FAITH =‘sit * sip * b. /sit/ sit — *! sip *

In this case, the role that Fl plays is similar to that of a positional faithfulness 22constraint (Beckman 1998) in that it allows privileged domains to be protected against markedness constraints. In (6a), [sit] is allowed to surface because F1, which is indexed to this morpheme in the lexicon (indicated by the subscript “1” on the underlying form), demands that it surface identical to its underlying form. (6b) demonstrates the need for the ranking OCP-coR>> FAITH, since in forms that are not lexically indexed to Fl the OCP effect is enforced. So in order to get the OCP effect at all, OCP-coR must outrank FAITH (6b), and in order to get some forms that violate the OCP, Fl must outrank OCP-coR. This approach essentially relies on a highly stratified lexicon, where words are indexed to whatever faithfulness constraints are responsible from protecting them from a given OCP effect. This is illustrated in (7), where various OCP-coR constraints are ranked with respect to each other, and lexically indexed faithfulness constraints are interleaved.

(7) F3 (protection from all OCP-coR effects) OCP-c0R[-son] [cLcont] F2 OCP-c0R) I (protection from OCP-COR[-son], OCP-C0R[-son] Fl (protection from OCP-coR) I OCP-coR

FAITH

Under this view, different portions of the lexicon are divided up among different “strata”. This general approach is unappealing, because it requires some additional mechanism to “assign” any given lexical item to a particular stratum of the lexicon (the number of which depends on how many clusters of lexical items are indexed so as to be protected from some

22 Including the notion of root vs. affix faithfulness, which is probably a closer analogue than other types of positional faithfulness constraints, such asthose defined interms of prosodic structure.

56 number of OCP constraints). Put more precisely, in order for this approach to work properly, it must meet some crucial conditions. If the rich (hypothetical) base has a completely even distribution of phonotactic structures, and also an even distribution of indexation options (orthogonally to phonotactic considerations), then this approach accomplishes what it’s supposed to do, which is generate an output set that replicates the frequency distribution of the lexicon. But unless all of these conditions are met, the approach will not derive the right results. One problem associated with these conditions is the mapping of output to input forms. By lexicon optimization, the learner should be taking output forms as input forms (especially since there are no alternations involved in the phenomena at hand). This calls into question the lexical indexation approach, and it also questions the learnability of a grammar based on lexically-indexed constraints.

3.2.3. Weighted Constraints

The second possibility being considered here is to assume that what the grammar should be modeling, or deriving, is not necessarily the lexical distribution pattern as such, but rather, the (gradient) wellfomedness judgments of speakers. This rests on the further assumption that wellformedness judgments (which are only hypothetical in the Gitksan case, but which are supported cross-linguistically by a wealth of empirical evidence) are themselves reflective of the gradient phonotactic patterns that exist in the lexicon. A way of modeling this relationship is with weighted constraints (Coetzee & Pater 2008). Rather than the strict domination of constraints found in OT, theories such as Harmonic Grammar employ numerically weighted constraints. Constraint violations are additive; the weighted sum of a candidate’s constraint violations constitutes a harmony score. Constraint violations are construed as negative numbers, and the candidate with the highest harmony score is selected as the output. Harmony scores translate into Acceptability scores, which can be compared across tableaux. This is how weighted constraints model gradient phonotactics: winning candidates with higher Acceptability scores are predicted to be more attested in the lexicon than winning candidates (from other tableaux) with lower Acceptability scores. In this way the weights of the constraints (and their violations) contribute toward a profile of the patterns found in the lexicon.

57 The weighted-constraints approach also has the added advantage of coming with a learning algorithm, which helps to account for the appropriate constraint weighting schema that gets acquired. This is perhaps one of the weakest aspects of the lexical-indexation model, and perhaps one of the strongest points of the weighted-constraints model. Furthermore, the learning algorithm doesn’t tap directly into any kind of calculation of frequency distributions as such, but is more indirect, as it takes into account observed lexical items, re-weights constraints based on those items, and those constraints in turn contribute to the relative acceptability of a given set of candidates. Because of the strengths of this model, it will be adopted in this thesis. An outline of how a grammar with weighted constraints works is given in section 3.3.1. Next, an approach will be laid out within a framework with numerically weighted constraints: Harmonic Grammar.

3.3. Harmonic Grammar

The theoretical framework that will be adopted to account for these gradient phonotactic effects in Gitksan is Harmonic Grammar (Legendre et al. 1990, 2006, Pater et al. 2007).

3.3.1. Overview

Following Coetzee & Pater (2008), the motivation for this framework is primarily that it captures gradient well- or ill-formedness distinctions between forms. In most theories of generative phonology, including OT, there is a firm distinction between forms that are grammatical, and those that are ungrammatical. There is, however, no differentiation in degree between grammatical forms, and likewise for ungrammatical forms. There exists, however, an extensive literature that suggests that speakers actually do differentiate between degrees of each. For instance, with respect to consonant clusters, many have noted that there are degrees of well-formedness amongst occurring and non-occurring forms (Chomsky & Halle 1965, Scholes 1966, Pertz & Bever 1975, Albright 2007). Berent et al. (2007) have discussed different degrees of ungrammaticality; that is, degrees of acceptability among words that are ungrammatical. A list of other works have done detailed psycholinguistic studies investigating the relative acceptability among well-formed words. For English these

58 include Bailey & Hahn (2001), Coetzee (2004, 2008), Greenberg & Jenkins (1964), Hammond (2004), Treiman et al. (2000), etc., and for Arabic Frisch & Zawaydeh (2001). One major problem with an OT approach with lexically-indexed constraints discussed in section 3.2.2 (which essentially protects certain lexical forms from OCP effects; cf. Coetzee & Pater 2006) that was pointed out by Hayes & Wilson (2008) is that while it models gradient phonotactics, it has no way of modeling gradient ill-formedness. Thus, there is no way to capture the fact that some non-occurring sequences are judged worse than others. Chomsky & Halle (1965) noted this type of effect in English, whereby words like brick are existing words, and non-existing words like buck, which are accidental gaps, are also judged well-formed. What is important about these judgments is that non-occurring forms like bnick are judged as better than other non-occurring forms like bdick. Both bnick and bdick contain ill-formed onset clusters, but experiments reveal that speakers judge the former to be better than the latter (Chomsky & Halle 1965, Albright 2007ab). Based on this kind of evidence, a phonological grammar should be capable of modeling this type of behavior, since this cannot simply be a matter of lexical frequency. An analysis of the OCP in Muna is provided by Coetzee & Pater (2008), who adopt an approach based in Harmonic Grammar (Legendre et al. 1990, 2006, Smolensky & Legendre 2006). Harmonic Grammar (HG) is a close relative to Optimality Theory (Prince & Smolensky 1993 [2004]), but differs in significant ways. One of the biggest differences is in the way that constraints are “ranked”. In OT, constraints are ranked in a strict domination hierarchy; in HG constraints are numerically weighted, and thus not in strict domination. Under standard OT, highly ranked constraints can be responsible for outcomes, and no amount of violations of a lower ranked constraint can trump the violation of a higher constraint. Under HG, instead of fixed rankings, constraints are assigned numerical weights, and these weights are multiplied by the constraint violations to yield a harmony score. Harmony is formally defined in (8):

(8) H(R) = 1CW(R) + 2CW(R) + 3CW(R) + ... WC(R) The notation for this formula is as follows: R is a given representation, 1C(R), 2C(R), etc. are the scores that R receives for each constraint (i.e. the number of constraint violations,

59 in OT), and ,1W ,2W etc. are the weights for each constraint (see also Coetzee & Pater 2008). Thus: Harmony equals the sum of the product of each score by each constraint weight for a given representation. The output form has maximal harmony, and intermediate harmony scores are taken to represent intermediately ill-formed candidates. Thus, the Harmonic Grammar approach addresses the criticisms of Hayes & 23Wilson (2008). Following Coetzee & Pater, violation marks in OT are converted to negative integer scores, and constraint weights are restricted to non-negative real numbers. It is also worth noting that while using numerical weights is a way of expressing constraint strength, this does not imply that two constraints cannot have the same weight (contra OT). There is nothing in the theory that prevents this state of affairs. This architecture is illustrated in (9), where the weightings for each constraint appear in the cell above the constraint, violations are listed as negative numbers, and harmony scores are located in the rightmost column.

(9) Constraints with weights Weight 2 1 H Input-i Constraint 1 Constraint 2 Outputl-l -l -1 Output 1-2 -1 -2

One arguably positive aspect of this type of architecture is that it allows for cumulativity, or “gang effects” (cf. Jager & Rosenbach 2006 for different varieties of these types of effect, as well as the implications they have for various models of grammar). This means that several violations of constraints with lower weights can overturn the effect of a single constraint with a higher weight. This type of effect is not possible under standard OT, where strict domination prevents any lower ranked constraints from having an effect on an output form. Of course, however, cumulative effects are possible in a theory of OT which allows for constraint conjunction; this will be discussed further in chapter 4. A cumulative

effect is illustrated in (10), where 3 violations of Constraint 2, which only has a weight of 1,

23 Hayes & Wilson’s (2008) own approach employs a Maximum Entropy grammar, which is also based on numerically weighted constraints and is a stochastic type of Harmonic Grammar.

60 “gang up” on Constraint 1, which has a higher weight (w = 2), but only a single violation. Thus, the output that violates Constraint 1, but satisfies Constraint 2 is the .24winner (10) Constraint weightings Weight 2 1 H

Input-2 Constraint 1 Constraint 2 Output 2-1 -3 -3 Output2-2 -1 -2

These types of cumulative effect have been noted for both syntax and phonology. The ganging up of markedness constraints in phonology has been noted by Pater (2008b) and Farris-Trimble (2008), and the ganging up of faithfulness constraints has been discussed at length in Farris-Trimble (2008). Under standard formulations of OT, strict domination would require violations of higher ranked constraints to trump violations of any lower ranking constraints. Thus, any violation of a higher ranking Constraint 1 will be worse than any number of violations of Constraint 2. Once a candidate is eliminated by means of constraint violation, there is no possibility of lower constraints having an effect on the outcome. This is illustrated in (11), which is (10) converted into a strict domination setting. This configuration “incorrectly” (in this hypothetical, schematic scenario) selects the first candidate as winner.

(11) Strict domination fails

Input-2 Constraint 1 Constraint 2 Output2-1 *** Output 2-2

A more interesting (and potentially more relevant to the case at hand) scenario is where there are two or more distinct lower-weighted constraints that gang up to override a single higher-weighted constraint. This can be considered in the context of the constraints responsible for the OCP effects, where the combined effects of constraints like OCP-coR, OCP-COR[-son]and OCP-coR[-son, acont] could potentially override IDENT[place],yielding an output form that is not identical to the input form (discussion of the theory of what OCP

24 Constraint weights were checked by the HaLP software (Harmonic Grammar with Linear Programming; Potts et al. 2007), which finds the minimal consistent weightings for the data.

61 constraints are employed is in the next section). This could happen even if all three of the individual weights for the constraints were considerably less than the weight for IDENT[place]. If the combined sum of the weights for these OCP constraints equals more than the weight for IDENT[place],then [d-ts] could be ruled out.

(12) Constraint weightings (hypothetical)

Weight 4 - 3 2 1 Input: /d-ts/ IDENT[placej OCPCOR[-son,ccont] OCPc0R[-sonj OCPc0R H d-ts -1 -1 -1 -6 d-p -1 -4

Again, under assumptions of strict domination, the violation of a higher ranked constraint cannot be overpowered by any number of lower-ranked constraints:

(13) Strict domination fails Input: /d-ts/ IDENT[placej OCPc0R OCPCOR[-son] OCPc0R[-son, (xcofl d-ts * * * d-p *!

As Legendre et al. (2006) point out, there could in principle be a state of affairs whereby the weights of each constraint conspired to approximate strict domination. In other words, the weightings of each constraint would be sufficient to prevent any type of overpowering effect from any constraints further down the hierarchy. Legendre et al. note, however, that this would require exponential constraint weighting.

“Suppose first that each constraint can be violated oniy once per candidate. What strengthswill yield the strict dominationhierarchy1C,>> ... >> 2C>> 1C >> ?0C To set the arbitrary origin of weights, let the strength of 0C be 1. Then obviously the weight of 1C must be greater than 1; to set the arbitrary scale of weights, let the weight of 1C be 2. Now in order that 2C have strict priority over both lower-ranked constraints,the cost of a violationof 2C must be higherthan the cost of violatingboth 1C and ,0C i.e., greater than 2+1=3; we can keep integer weights by taking the strength of 2C to be 4. Continuingthis logic,we see that the strict dominationcan be achieved by setting the weight of Ckto be 2k We will refer to this as exponential weighting: the strengths of constraints must grow exponentially as the hierarchy is mounted.”(910)

Thus, there is the possibility that there are areas of the grammar whereby the constraints behave as they would under strict domination. It is important to note the

62 relationship here between weighted and ranked constraints, as there are claims that the two systems can co-exist within the phonological component (Smolensky & Legendre 2006).

3.3.2. Constraints

Based on the results of chapter 2, the relevant constraints for the OCP[place] effect in the Gitksan data include the following:

(14) OCP[place] constraints OCP-LAB No sequences of labials OCP-COR No sequences of coronals OCP-D0R No sequences of dorsals

These constraints penalize cooccurrences of the features [lab], [cor], [don; generally speaking, autosegments on a particular tier. Given that there are cases of consonant-vowel interaction in the language (e.g. labial assimilation, guttural lowering), it is understood here that, assuming the feature geometry of Clements & flume (1995), there is a distinction between C-place and V-place instances of these features. Following Suzuki (1998) with regard to OCP effects over varying distances, these constraints penalize consonants that are non-string adjacent (see also Coetzee & Pater 2008:21). The constraints in (14) apply to sequences involving a consonant and the next following consonant (where intervening vowels are ignored). In order to derive the difference between transvocalic and long-distance effects (see examples 11 vs. 15 in chapter 2), a separate set of constraints will be introduced below based on this same formulation. In order to capture the subsidiary effects due to other contributing features, the following constraints on coronals are necessary. The set of such constraints is motivated by the empirical findings of chapter 2 and from other languages, where classes like coronal show evidence of subdivisions into smaller classes based on major class and manner features. Thus, OCP-coR is accompanied by OCP-COR[+son], OCP-COR[-son], and OCP-coR[-son] [ctcont]. In order to make OCP-coR, OCP-D0R and OCP-LABcomparable (meaning each of these constraints is the sum of all possible gang effects for each place of

63 articulation), then each of the constraints in (14) must be similarly broken down into relativized versions:

(15) Other contributors OCP-LAB[+son] No adjacent labial sonorants OCP-LAB[-son] No adjacent labial obstruents OCP-LAB[-son] [cLcont] No adjacent labial obstruents agreeing in continuancy

OCP-coR[+son] No adjacent coronal sonorants OCP-C0R[-son] No adjacent coronal obstruents OCP-coR [-son] [ctcont] No adjacent coronal obstruents agreeing in continuancy

OCP-D0R[+son] No adjacent dorsal sonorants OCP-DOR[-Sofl] No adjacent dorsal obstruents OCP-DOR[-son] [acont] No adjacent dorsal obstruents agreeing in continuancy

While the resulting set of constraints is uninteresting for some classes of consonants (e.g. there are no dorsal sonorants in the language, so OCP-DOR[+son]is completely inert), it is still necessary to assume a relatively unbiased set of constraints, and then allow the learning procedure to arrange those constraints into a ranking (i.e. a set of weightings). These constraints are in conflict with the faithfulness constraint that demands identity between input and output correspondents, IDENT (McCarthy & Prince 1995):

(16) IDENTIO-PLAcE(ID-PLAcE): Let a be a segment in the input, and 13be any

correspondent of a in the output. If a is [yplace], then 13is [yplace]. (Correspondent segments are identical in place features)

The interaction of these constraints is most easily illustrated with the labial and dorsal pairs, respectively, and using only the basic OCP[placej constraints (where P = any labial consonant, K = any dorsal consonant, T any coronal plosive, L = any coronal sonorant, and S = any coronal fricative).

(17) Tableaux for labial-labial and dorsal-dorsal sequences Weight ..4 I I± I H Input: P-P ID-PLACE I OCP-LAB J OCP-C0RJ OCP-D0R P-P -1 —1

. P-T —1 —4 I

64 Weight 4 1 1 1 H Input: K-K ID-PLACE OCP-LAB OCP-C0R OCP-D0R K-K -1 -1 K-T -1 -4

Since all of these outputs constitute acceptable forms for lexical items (i.e. there are no combinations of place features alone that are banned outright), the combined weights of the OCP constraints must be lower than the weight for ID-PLACE. Coetzee & Pater (2008) correlate the OlE values from the Muna data not directly with the harmony scores, but with an acceptability score. This score is derived by taking a candidate’s harmony score and subtracting the harmony score of its closest (most harmonic) competitor (Acceptability(x) = H(x) — H(y)). In this way, structures that are well-formed receive a positive acceptability score, whereas structures that are ill-formed receive a negative acceptability score. Among each type (positive vs. negative), a higher acceptability score is intended to translate into greater acceptability, and this is how candidates from different tableaux (i.e. from different inputs) can be compared in a meaningful way. This is how HG accounts for gradience in both well-formed and ill-formed structures. The role that acceptability plays can be seen in (18), where (winning) candidates across tableaux can be compared in terms of their acceptability scores. Since the well-formed outputs T-L, T-TS, and T-T each receive an acceptability score of 3, 2, and 1, respectively, this translates into the statement that T-L is “better” than T-S, and T-S is “better” than T-T.

(18) Tableaux for coronal-coronal sequences Weight 4 1 1 1 HA Input: T-L ID-PLACE OCP-C0R OCP-C0R[-son] OCP-C0R[-son, acontj — T-L -1 -1 3 T-P -1 -4

Weight 4 1 1 1 HA Input: T-S ID-PLACE OCP-C0R OCP-C0R[-sonj OCP-C0R[-son, xcontj — — T-S -l -l -2 2 T-P -1 -4

65 Weight 4 1 1 1 HA Input: T-T ID-PLACE OCP-C0R OCP-C0R[-son] OCP-C0R[-son, acont] T-T -1 -1 -1 -3 1 T-P -1 -4

For sequences of consonants that are heterorganic, the only relevant constraint is ID-PLACE, which will ensure that output forms are faithful to their inputs. Again, the acceptability scores for the outputs in (18) can be compared with the scores in (19), where T-K proves to be “better” than any of the candidates in (18), including T-L.

(19) Heterorganic pairs are not penalized Weight 4 1 1 1 HA Input: T-K ID-PLACE OCP-C0R OCP-C0R[-son] OCP-C0R[-son, acont] — T-K 04 T-P -1 -4 -4 T-T -1 -1 -1 -1 -7 -7 T-L -1 -1 -5 -5

Since there are no cases of absolutely banned cooccurring consonants (independent of other phonotactic considerations), ID-PLACE will be more highly weighted than the summed weights of all of the constraints that are available to gang up on ID-PLACE. In other words, all lexical items will surface identical to their inputs. With respect to the general learning algorithm to be discussed below, it is assumed (by lexicon optimization) that input forms posited by the learner are identical to the output forms that are encountered by that learner. By encountering the output forms of a language, the learning procedure induces a grammar; in other words, the learner induces a set of weights for the constraints. These points will be discussed shortly in the following section.

3.3.3. Learning the Grammar

In order to see how constraint weights and acceptability link to the empirical data of chapter 2, a specific learning procedure must be assumed. The success of the learning procedure needs to be measured by examining the acceptability scores that are generated from the constraint weightings, and then comparing these to the empirical lexical data (i.e. the OlE values) to determine whether there is a correlation. Thus, while there is at present nothing

66 meaningful to compare acceptability scores to in Gitksan (i.e. psycholinguistic judgment tasks), this approach assumes that there is some link or correlation between the grammar (i.e. constraint weightings) and the lexicon (i.e. OlE values). The learning algorithm employed is a modified version of the Gradual Learning Algorithm (Boersma 1998, Boersma & Hayes 1998). This algorithm captures the fact that learning is gradual, and occurs item-by-item. Following Coetzee & Pater (2008) and Pater (2008a), it is assumed here that the learning algorithm is error-driven, it is on-line, and consists of an update rule (what Coetzee & Pater refer to as a weight update rule, or stochastic gradient ascent). In this model, the learner is supplied with outputs, and the grammar’s current state (the current weightings) is used to produce an output form. The output that is presented to the learner (the learning datum) is assumed to be faithful to the input to the grammar since there are no alternations (cf. Moreton 1996 on input-output mapping; see also Hayes 2004 and Prince & Tesar 2004 on this assumption in theories of phonotactic learning). When the output that a learner produces is different from the learning datum, learning is triggered and the relevant constraints are adjusted by a given amount (n). The update rule is then employed to bring the grammar into a state more consistent with the learning data. In Coetzee & Pater (2008), the update rule is as follows:

(20) Update rule for HG Add n*(vE — vC) to each constraint weight

Where yE is the number of violations incurred by the error, and vC is the number of

violations incurred by the correct form, and 0

67 value for n, which is equivalent to the plasticity value in the Gradual Learning Algorithm and

determines the learning rate, is set here at 1 since we’re dealing with the initial phase of learning (see Coetzee & Pater 2008, Pater to appear). If n is set higher, then learning proceeds in faster, more coarse-grained jumps than if n is set lower. The pointing hand picks out the winner selected by the grammar (the error), while the check mark indicates the correct surface form. Arrows indicate the direction of constraint re-weighting. (The forms used here are genuine lexical items in Gitksan):

(21) Constraint re-weighting Weight 100—> 100—> —50 H Input: /lanJ ‘fish eggs’ OCP-coR[+sonj OCP-coR ID-PLACE “1-n -1 -1 -200 ‘l-m -1 -50

As illustrated in (21), learning is triggered by the fact that the error deviates from the correct form. In this case, the weighting of the constraints that are violated by the correct form, namely OCP-CORand OCP-COR[+son]are “nudged down”, or have their weights decreased. The constraint violated by the error, ID-PLACE, is “nudged up”, or has its weight increased. The need for both weight increases and decreases (rather than just decreasing, which is essentially the mechanism employed by Tesar & Smolensky 2000) has been argued for by Boersma (1998), and Boersma & Pater (2008).25 If (21) constitutes a hypothetical first token in the learning process, then the weights for each constraint afterward would be 99 for each of the relevant markedness constraints and 51 for ID-PLACE. (22) illustrates the state of the grammar before and after this single datum is encountered:

(22) Weights before and after learning

• Constraint - Initial Weight Weight after learning

I OCP-coR[+sonj 100 99 I OCP-COR 100 99 ID-PLACE 50 51

25 While there is a proof of convergence for constraint demotion in OT (Tesar & Smolensky 1998, 2000), there is not an explicit proof for a demotion-only approach in HG. There are, however, proofs available for the promotion/demotion approach in HG. Thanks to Joe Pater for pointing this out to me.

68 It might now be informative to observe the learning procedure after several lexical items are encountered. (23) illustrates how all of the relevant OCP-related constraints (including ID- PLACE) react to gradual learning, starting from the initial state:

(23) All constraints, tableaux ordered in learning time Datum 1: Weight 100 100—> 100 100 100 ÷—50 H Input: /k’e:q/ OCP- OCP- OCP- OCP-coR OCP-coR ID- ‘run away, LAB DOR COR [-son] [son, acont] PLACE flee’ v7k’-q -1 -100 k’-t -1 -50

Datum 2: Weight 100 100—> 100 100 99 —51 H Input: /sil/ ‘drunk, OCP- OCP- OCP-coR OCP-coR OCP- ID- intoxicated’ LAB COR [-son] [-son, acont] DOR PLACE s-1 -1 -100 s-t -1 -1 -1 -251 s-s -1 -1 -1 -1 -351 s-m -1 -51

Datum 3: Weight 100 100—> 100—> 99 99—> ÷—52 H Input: his! OCP- OCP-C0R OCP-CoR OCP- OCP-C0R ID- ‘finish’ LAB [-son] [-son, cLcont] DOR PLACE +-s -1 -1 -1 -299 +t -1 -1 -1 -251 1-1 -1 -1 -151 +-m -1 -52

Datum 4: Weight 100 99 99 99 98—> —53 H Input: hit! OCP- OCP- OCP-C0R OCP-D0R OCP- ID ‘wedge’ LAB COR[-son] [-son, ctcont] COR PLACE V1-t -1 -98 1-s -1 -98 1-1 -1 -98 l-m -1 -53

69 Datum 5: Weight 100 99—> 99 99 97—> —54 H Input: /t’a:li ‘to OCP- OCP- OCP-coR OCP- OCP- ID- pick (berries)’ LAB COR[-SOfl] [-son, Lcont] DOR COR PLACE Vt’aaj -1 -1 -196 t’aal -1 -97 t’aap -1 -54

The effect of each stage of re-weighting can be seen in (24), where the weight for each constraint is given for each stage of learning (i.e. after each datum is encountered):

(24) Constraint weights at each state of learning

Constraint Initial State Datum 1 Datum 2 Datum 3 Datum 4 Datum 5 OCP-LAB 100 100 100 100 100 100 OCP-D0R 100 99 99 99 99 99 OCP-C0R[-son, acont] 100 100 100 99 99 99 OCP-COR[-son] 100 100 100 99 99 98 OCP-COR 100 100 99 98 97 96 ID-PLACE 50 51 52 53 54 55

At some point the grammar reaches an equilibrium state, or convergence point. Convergence is when the grammar consistently yields the correct output for every single lexical item that it encounters. Once convergence is reached, there is no further learning (in this case, adjustment of constraint weights). This means that OCP constraints will be “nudged down” in weight, and ID-PLACE will be “nudged up” in weight until every single lexical item, or at least every single attested type of C-C combination has been encountered and surfaces faithfiilly. Since there is no type of C-C combination (at least in CVC sequences) which is unattested, or categorically absent, then this means that convergence will be reached when all of the OCP[placej constraints have weights that are less than that of ID PLACE. With regard to the “ganging-up” problem mentioned in (15), this also specifically means that any potential gangs of OCP constraints must end up with a sum that is less than the weight for ID-PLACE. Take for example the set of OCP constraints {OCP-C0R, OCP COR[-sonj, OCP-C0R[-son, xcontj}. These constraints can each be violated simultaneously by a single candidate such as [d-ts], or even the word /1is/ ‘finish’, which was used as Datum

3 in (23) above. Since this is the case, the sum of all of the weights for these constraints must

70 be lower than the weight for ID-PLACE, otherwise an output like [d-ts] would fail to surface. This is a direct consequence of the mechanics of the learning algorithm: convergence is where no attested C-C combination will trigger a repair (this holds true for novel items, as well, such as borrowings or nonce words, though this has yet to be tested for Gitksan). Since many of the OCP constraints exist in a subset relationship, there must be some way that learning will continue even though certain constraints are no longer being affected by the learning process. An example helps to illustrate: A more specific constraint like OCP COR[-son,cLcont]will end up being nudged down in weight slower than its more general (or superset) counterparts, such as OCP-coR. This is because there are more forms in the lexicon that violate OCP-coR than violate OCP-coR[-son, acont]. Thus, by the time OCP COR[-son,contJ is nudged down below ID-PLACE, OCP-coR will be much lower in weight. However, at this point, there will be no further adjustments to these constraint weights (i.e. no more learning) for coronal-coronal sequences of any kind, even though adjustments may still be ongoing for labial-labial pairs (or any other pairs that are very rare). An immediate question is, how does ID-PLACE continue to rise in weight? The answer is because even though there are no further adjustments to OCP-coR, ID-PLACE will continue to be nudged up in weight due to errors resulting from labial-labial or dorsal-dorsal inputs. ID-PLACE and OCP-coR are not in a direct trading relationship, and that is because ID-PLACE is defined over all places. Thus, labial-labial sequences will violate ID-PLACE, as will dorsal-dorsal sequences, meaning ID-PLACE will continue to be nudged up in weight even though OCP COR is no longer being nudged down. In the same fashion, OCP-LAB will continue to be nudged down, but by the time it is overtaken by ID-PLACE and no longer continues to nudge downward in weight, it will still be higher than OCP-coR. This is because ID-PLACE has been rising in weight to meet OCP-LAB.

At convergence the weights of each of the constraints are set. This approach to learning amounts to viewing the constraint rankings (or in this case weightings) as matching the frequency profile of the lexicon in the end state. Following Coetzee & Pater (2008), the type of learning algorithm discussed here will by definition lead to often-violated (gangs of) OCP constraints (such as gangs of OCP-coR) that end up with lower weights than seldom- violated constraints (such as OCP-LAB). It is in this sense in which the grammar reflects the statistical patterns attested in the lexicon of the language. As can be seen in the learning

71 process (23) or simply the course of weight adjustment (24), incoming lexical items “shape” the grammar at each step. Readjustments to constraint weights are undertaken each time the grammar fails to produce the correct output form, and this process takes place until the grammar consistently produces correct outputs. Thus by convergence, multitudes of lexical items have contributed in shaping the grammar, giving it a profile that reflects the lexicon. The actual application of the learning process described here uses the Gradual Learning Algorithm implementation in Praat (Boersma & Weenink 2007). In order to employ Harmonic Grammar (rather than Stochastic OT), the Linear OT evaluation mode was selected (see also Coetezee & Pater 2008). This setting ensures that positive real numbers are used (as opposed to the ‘Harmonic Grammar’ setting, which allows for negative weights). This results in the update rule being applied. Since variation is not observed with respect to OCP-related phenomena, the setting for noisy evaluation was set to 0. Other settings were the default settings in Praat. Learning data included all observed sequences of non-identical homorganic consonants from Gitksan. In other words, the frequency of each homorganic pair across the lexicon was used as the learning distribution. The constraint weights resulting from the application of the learning algorithm are presented in (25). OCP-PHAR, which will be dealt with in chapter 5, has been included here for completeness. Notice that ID-PLACE is weighted the highest, at 185.84. The extreme weighting of this constraint ensures a faithful input-output mapping.

72 (25) Constraint weights after application of learning algorithm

- Constraint Weight after learning ID-PLACE 185.8 OCP-D0R[+son] 100.0 OCP-LAB[-son] 100.0 OCP-LAB[-son,acont] 100.0 OCP-PHAR 97.1 OCP-LAB[+son] 90.2 OCP-LAB 89.2 OCP-COR[+son] 87.6 OCP-COR[-son,xcont] 83.2 OCP-DOR[-son,acontj 73.8 OCP-COR[-son] 60.5 OCP-D0R[-son] 51.9 OCP-DOR 46.6 OCP-coR 31.0

The fact that OCP-DOR[+son],OCP-LAB[-son] and OCP-LAB[-son,cLcont]all have weights at 100 reflects the fact that no such pairs were encountered by the learner (because these pairs are absent in the lexicon), and thus the weight for these constraints remained the same as it was in the initial state. This is unsurprising for pairs of dorsal sonorants (which would violate OCP-D0R[+son]) and pairs of non-identical labial obstruents agreeing in continuancy (i.e. no /p-p’/ pairs): these consonant types do not exist in Gitksan, and hence pairs of these consonants would not be encountered by a learner. The fact that OCP-LAB[ son] remains at 100 reflects a categorical restriction on labial obstruents cooccurring. It is also worthwhile to note that OCP-PHAR, with a weight of 97.1, is very close to 100, indicating that very few pairs of guttural consonants were encountered by the learner. In active, synchronic terms, the phonological grammar is doing nothing with respect to consonant cooccurrence effects; faithfulness rules all lexical fonns, and any lower weighted, “subterranean” OCP constraint has no visible effect. These constraints are weighted too low to trigger a repair (i.e. dissimilation). Output forms surface faithful to input forms. In very simple terms, the grammar is mirroring the lexicon: the grammar is a device used in acquisition which tracks lexical statistics, but plays no active role in determining output forms, per se. It is only when psycholinguistic tasks are presented to speakers (as in psycholinguistic judgment tasks) that the effect of the OCP constraints becomes visible.

73 With the weights from the learning simulation in hand, the acceptability scores for various consonant pairs can be compared with their OlE ratios. The idea is that the more attested a consonant pair is in the lexicon (i.e. the higher the O/E ratio), the higher the acceptability score. If the constraints are assigned weights that indirectly correspond to OlE values (by means of an acceptability score), then the model can begin to describe how gradient wellformedness is predicted by the grammar. What results is this: across these different tableaux, the winning candidates differ with respect to their acceptability scores. This is due to the fact that each winning candidate (which matches the input) will violate different constraints, resulting in different harmony scores. The harmony scores, in turn, are converted into acceptability scores by taking the harmony score of a given candidate and subtracting the harmony score of the most harmonic competitor. This results in acceptability scores that are positive for winning candidates, and negative scores for losing candidates (only consonant pairs are considered here, rather than real lexical items):

(26) Gradient wellformedness Weight 185.8 100 100 90.2 89.2 H A Input: ID-PLACE OCP-LAB OCP-LAB OCP-LAB ocP p-rn [-soni [-son, ctcontl [+son] LAB p-m —1 -100 85.8 p-n —1 -185.8

Weight 185.8 90.2 100 51.9 46.6 H A Input: ID-PLACE OCP-D0R OCP-D0R OCP-D0R OcP k-x [+son] [-son, ctcont] [-son] DOR

k-x —1 —1 -98.5 87.34 k-t —1 -185.8

Weight 185.8 87.6 83.2 60.5 31 H A Input: ID-PLACE OCP-coR OCP-coR OCP-coR OcP t—l [+son] [-son, acontl [-soni COR t-l —1 -31 154.8 t-k —1 -185.8

Weight 185.8 87.6 83.2 60.5 31 H A Input: ID-PLACE OCP-C0R OCP-coR OCP-CoR OCP t-ts [+sonj [-son, ctcont] [-son] COR t-ts -1 -1 -1 -174.7 11.1 t-k -1 -185.8

74 Weight 185.8 89.2 31 H A Input: t-p ID-PLACE OCP-LAB OCP-coR t-p 0 185.8 t-k -l -185.8

Importantly, these acceptability scores can be compared across winning candidates from different tableaux. As in the example above, the acceptability score for [p-rn] is lower than that for [t-tsj, which in turn is lower than the score for [t-l], which in turn is lower than that of [t-pJ. This gives the effect of reflecting the lexical patterns of the language: [p-rn] is relatively more underrepresented than [t-ts], which is itself more underrepresented than [t-l], etc. Heterorganic consonant pairs like [t-p], which exhibit no OCP effect, will receive the highest acceptability score.

(27) Acceptability rating by consonant pair (by acceptability) Consonant Pair Acceptability Relevant OCP constraint violated OlE t-ts 11.1 OCP-COR[-sonacont], OCP-coR[-sonj, 0.28 OCP-coR p-rn 85.8 OCP-LAB 0.16 k-x 87.34 OCP-D0R,OCP-DOR[-sonJ 0.54 t-l 154.8 OCP-coR 0.75 t-p 185.8 None 1.21

It is interesting to note that not all acceptability-attestedness comparisons came out as predicted by the model. Since labial-labial pairs have an OlE of 0.16, then they are predicted to be less acceptable than t-ts pairs, which have an OlE of 0.28. Under the current constraint weightings, this is not the case; t-ts has a lower acceptability rating (11.1) than p-m (85.8). All other consonant pairs, however, behaved as predicted with respect to acceptability and attestedness. Finally, differences between the final constraint weightings in Gitksan and other languages which do not exhibit the same OCP effects should be discussed. Earlier it was hypothesized that in languages like Gitksan, knowledge of the OCP patterns (in the form of graded acceptability scores/judgments) emerges during the course of learning as a result of OCP constraints being nudged down in weight at different rates, while ID-PLACE was continually nudged up. This produced a constraint weighting that reflected the gradient

75 patterns in the lexicon. With this in mind, languages without OCP effects would have to follow a different path during learning. A language without significant effects in the lexicon would mean that the learner would be encountering numerous OCP-violating forms, presumably at chance occurrence, and presumably also with equal probability as OCP respecting forms. Learning would nudge the OCP constraints so low in the end state that they would make a negligible contribution to the calculation of harmony scores, and hence also to acceptability scores (though interestingly the OCP constraints should still have some effect, as weighted constraints are additive). The opposite would characterize a language with categorical OCP effects: in this type of language, the absence of OCP-violating forms encountered by the learner would keep the OCP constraints at the same weight as in the initial state, with no adjustments in weight during learning.

3.4. Distance in OCP Effects

With the learning algorithm now in place, it is worthwhile to discuss the role of distance in the OCP effects. As noted above, the OCP effects are strongest when there is less distance between two consonants. The strength of the effect then diminishes as more phonological material intervenes. Following Suzuki (1998) these constraints are defined as applying to sequences of offending structures separated by one ore more consonants.

(28) OCP[place] constraints -LABOCP-LAB-C No long-distance labial-labial pairs -coROCP-coR-C No long-distance coronal-coronal pairs 0-D0ROCP-D0R-C No long-distance dorsal-dorsal pairs The precise definitions of these constraints involve the 0C notation, which if taken literally, means zero or more consonants. Thus, these constraints will penalize both the long-distance cases, as well as the “local” (transvocalic and adjacency) cases. This is a desirable outcome, as it sets the long-distance constraints into a subset, or stringency relation with the transvocalic constraints, which in turn means that these constraints will sink down in weight even lower than the generic OCP-coR constraint during the learning process. This runs parallel to constraints like OCP-coR and OCP-coR[-son, acontj being in a subset relation and their relationship in terms of weight adjustment during the course of learning.

76 With the idea in hand that the markedness constraints don’t actually contribute anything in terms of selecting the winning output candidate, it can still be seen how they contribute to an overall larger pattern in the lexicon. The tableaux below illustrate how the differences between transvocalic and long-distance OCP effects are derived by these different sets of constraints. The weights for OCP[distance] constraints will indeed be very low (again, using arbitrary numbers here), as there are no long-distance effects that are significant for OCP[place] effects. It is worth mentioning this difference in distance, though, as the notion of distance will be relevant for the assimilatory effects governing dorsals and glottalized consonants.

(29) Transvocalic effect Weight 100 30 20 10 H Input: /laqs/ ‘to ID-PLACE OCP.-coR OCP-COR[-sonj OCP-coR wash one’s body’ [-son, acont] ‘laqs -1 -10 laqx -l -100

(30) Long distance effect

Weight 100 30 20 10 1 H Input: /se:GalI ID-PLACE OCP-coR OCP-C0R OCP- -c0ROCP-c0R-C ‘be rough’ [-son, acontj [-son] COR ‘se:GQ1 0 -1 -l se:Gctp -l -100

As was noted with respect to the role that lower-weighted constraints play, the long-distance constraints do not interact per se with the transvocalic constraints; rather they accumulated their own constraint weights during the learning process. Since these constraints will be very often violated by lexical output forms (i.e. long-distance LAB-LAB pairs are not underrepresented, and are about as common as long-distance LAB-C0R or LAB-DORpairs), these constraints will be reduced in weight during the learning process to a value that will

have no effect in terms of lexical statistics (which is why the smallest positive integer 1 was chosen here). In this way the grammar can mirror both the transvocalic and long-distance effects by fixing the weights of each set of constraints, respectively.

77 3.5. Conclusion

This chapter has focused on the restricted nature of cooccurring consonant pairs in the Gitksan lexicon. In a study of roots in the language, it was shown that there is a moderate OCP[place] effect at play, where labials were most highly affected, then dorsals, then coronals. It was also shown that major class and manner features contributed to this effect. It was also shown that a gradient consonant harmony system exists in the Gitksan lexicon, whereby glottalized consonants tend to cooccur with glottalized consonants, velars tend to occur with velars, and uvulars tend to occur with uvulars. These effects were modeled using Harmonic Grammar and numerically weighted constraints. Highly weighted faithfulness constraints played the role of determining the correct output form, while lower-weighted OCP constraints were responsible for contributing towards a frequency profile of the lexicon. This approach is desirable in that it correctly accounts for the data, while at the same time constitutes a plausible learning scenario.

78 Chapter 4: Laryngeal Features

4.1. Introduction

As noted in earlier chapters, there are contrasting sets of pulmonic and glottalized obstruents, as well as sets of plain and glottalized sonorants in the Gitksan consonant inventory. Continuing from where chapter 2 left off, the gradient generalizations concerning glottalized segments over the Gitksan lexicon will be discussed. As outlined in chapter 2, there is a gradient agreement among pairs of glottalized consonants. It will be shown that this gradient agreement differs from similar types of agreement found cross-linguistically, and from the general OCP patterns of chapter 2 in that place features do not play a guiding role in the nature of agreement. The discussion then turns to the phonological status of the feature [voice], and how cooccurrence patterns lend themselves to an analysis of plosive voicing as an active phonological pattern. An analysis is then provided for the ejectives and fricatives, which resist this process of voicing. It is claimed that these segments are [constricted glottis] and [spread glottis], respectively, and that there is a general constraint on the association of multiple laryngeal features to a single obstruent (but not a single sonorant, due to the possibility of both glottalization and redundant voicing surfacing on these segment types). Like many languages of the Northwest Coast of North America, Gitksan has a rich set of glottalized consonants, including ejective stops and affricates and glottalized sonorants. Gitksan has the following underlying consonant inventory; the glottalized consonants have been bolded.

(1) Gitksan consonant inventory Obstruents p t Is k k’ q p’ t’ Is’ t+ k’ k”’ q’ s + x x x h

Resonants m w n 1 j Ifl’ W’ fl’ 1’ j’

79 There is a near symmetry with respect to the pulmonic and ejective plosives, as well as the plain and glottalized sonorants: each pulmonic plosive and plain sonorant is paired with a glottalized version. The affricate 1W! is unpaired, which is the only element making the

symmetry incomplete. There are more glottalized obstruents in this system than pulmonic oral obstruents; while typologically unusual, this is often found in the Pacific Northwest, where [t+’] appears without a pulmonic counterpart in many languages (Thompson &

Kinkade 1990). Not all ejectives in the language show up with the same frequency; /p’/ is much more limited in this sense, and can be considered marginal in the language. Voiced plosives, while not in the underlying consonant inventory, exist allophonically in the language. In Gitksan, there is an alternation between the voiced and voiceless stops and affricates. Pulmonic plosives in prevocalic position become voiced; in non-prevocalic (i.e. pre-consonantal or word-finally), these plosives are voiceless. This can be illustrated below, where (2a) shows an intransitive/transitive alternation, and (2b) illustrates a form suffixed with a pronominal affix.

(2) a. /kup/ [gupj ‘to eat (intrans)’ /kup-it/ [gu.bit] ‘he/she ate it’ eat-3SG

b./n-pip/ [ni.bip} ‘maternal uncle’ /n-pip-nJ [ni.bi.binj ‘your(sg.)maternal uncle’ PFX-uncle-2SG

While it has been claimed by Hoard (1978) that ejectives also undergo voicing (with the result being implosive consonants), this has been disconfirmed by instrumental studies performed by Ingram & Rigsby (1987) and Rigsby & Ingram (1990), who show that ejectives fail to undergo voicing altogether in the language. Impressionistic and instrumental data that I have collected on the language also indicates this. Plosive voicing, and the issues surrounding it, will be discussed further in section 4.3. Contrasts between plain and glottalized consonants are illustrated below. Contrasts between pulmonic plosives and ejectives are illustrated in (3), while plain and glottalized sonorants are contrasted in (4).

80 (3) Pulmonic vs. ejective stops [da:x] ‘circumference, outer circle’ [t’ax”’] ‘to sweep’ [ga+x’9 ‘to pierce, stab, stick’ [k’a:+] ‘aside, to one side’ [Ga:k”] ‘sinew’ [q’a:x] ‘wing, feather’ [a+] ‘to eat up (trans)’ [ts’a+X] ‘to laugh’ [doq] ‘to take, to pick up (p1)’ [doq’] ‘to be deaf [ha:waq] ‘birch tree’ [ha:naq’] ‘woman (p1)’

(4) Plain vs. glottalized sonorants [na:] ‘out of the woods (v. proclitic)’ [n’a:] ‘to complete an action’ [flaX] ‘snowshoe’ [n’aX] ‘bait’ [masx’9‘to be red ochre-coloured’ [m’asx”’J ‘to fart, to sting (intrans)’ [mats] ‘arrive, run in number (of salmon)’ [m’ats] ‘to hit, strike’ [muk”] ‘sawbill duck’ [m’uk”’] ‘to catch (fish)’ [wa] ‘name’ [w’aj ‘to find, to get to someplace’ [jim] ‘to smell (trans)’ [j’imq] ‘whiskers, beard’ [ts’al] ‘half-smoked salmon’ [lo’oba ts’al’] ‘fool hen’

Given these contrasts, the assumed underlying representations for plain (non-glottalized) vs. glottalized plosives in terms of laryngeal features is as in (5):

(5) a. Root b. Root

Laryngeal

[constricted glottis]

In (5a), there is only a root node with no specified laryngeal features. This is the (output) representation assumed for pulmonic obstruents (motivating evidence is presented in section 4.3.1). (5b) illustrates a glottalized consonant, where there is a laryngeal node present, and [cg] is specified as a laryngeal feature. Again, these are underlying representations, though assuming richness of the base (Prince & Smolensky 1993 [2004]) will also yield consistent results with the constraints proposed in 4.3. Once a pulmonic plosive undergoes allophonic voicing, it acquires the feature [voice] (as in 6) (which also entails that it acquires a laryngeal

81 node), whereas non-prevocalic pulmonic plosives remain without a laryngeal node in their surface representation. Since there exists no outright evidence for binarity, it is assumed here that these laryngeal features are privative. The claim here is not that there is a feature [-voice] which remains underspecified into the output, but rather, that there is no feature value [-voice] that exists to begin with. Evidence for [voice] being a privative feature is presented by Mester & Ito (1989), Cho (1990), and Lombardi (1995).

(6) Root

Laryngeal

[voice]

Sonorant voicing is often a more complex issue than obstruent voicing. Under some approaches, where sonorant voicing is viewed as the same type of phenomenon as obstruent voicing, voiced sonorants would have a representation as in (6), while glottalized sonorants would require a different representation and be more complex than their obstruent counterparts in (5b), as the Laryngeal node would dominate both [voice] and [cg]. These issues are discussed further in section 4.3. The results of chapter 2 indicated that there is a tendency for glottalized plosives to cooccur with other glottalized plosives in what was termed a gradient glottalization agreement. This gradient agreement will be analyzed in the following section.

4.2. Glottalization Agreement

Turning back to the results from chapter 2, we can see that within Gitksan roots, there is a gradient assimilatory tendency found within the glottalized consonants. Within CVC pairs, there exists a gradient laryngeal agreement in plosive-plosive pairs. This is illustrated in (7), where the OlE values for pulmonic-ejective (in either order), pulmonic-pulmonic, and ejective-ejective pairs are compared.

82 (7) OlE values for plosives CxvCv T-T’ / T’-T 0.75** 8/50.9) T-T 1.13** (56/49.5) T’-T’ 1.75** (15/8.5)

It was shown in chapter 2 that there is no evidence of an assimilatory effect for glottalization in sonorant-sonorant CVC pairs, nor was there an effect in obstruent-obstruent pairs which was independent of the specific plosive-plosive effect. It was also shown that the effect for plosives only holds for the transvocalic pairs, and not in long-distance contexts. It is now relevant to place this gradient laryngeal agreement into a broader typological context. There can first be drawn a dichotomy between assimilatory and dissimilatory tendencies that affect laryngeal features. In the Mayan language Chol (Coon & Gallagher 2007), roots may contain two ejective consonants, but only if the consonants are completely identical. This restriction is categorical (cf. Yip 1989, MacEachern 1999 for a general discussion of languages that display categorical behaviors such as Chol).

(8) Chol ejectives (Coon & Gallagher 2007) *k’ap’ v’k’ak’

This restriction can be considered a type of categorical dissimilatory effect. In other words, the permissibility of k’-k’ type sequences is an exception to what would otherwise be an exception to the effect. In the formalism of chapter 3, this effect would be derived by a constraint like OCP[cg], which prevents multiple occurrences of the feature [cg]. In order to derive the forms that do have multiple ejectives, the total identity exemption effect is appealed to (cf. MacEachern 1999, Frisch 2004). In MacEachem’s (1999) work on cooccurrence restrictions, a constraint that prohibits non-identical segments (BEIDENTICAL) is employed. Since cases of complete identity are not the focus of this thesis, it will be assumed that some version of this constraint is highly weighted in the grammar. The dissimilatory type of pattern found in Chol can be contrasted with assimilatory patterns that are typically found in the consonant harmony literature. In discussions of

83 consonant agreement, it has been noted by MacEachem (1999), Hansson (2001), and Rose & Walker (2004) that laryngeal agreement often is dependent on homorganicity; that is, if consonants share place of articulation, they must also agree in laryngeal 26features This type of pattern is found in Bolivian Aymara, Hausa, and Tzujutil: stops in these. languages are not required to match in terms of place features, but if they do happen to be homorganic, then the stops must also be identical in terms of laryngeal features. These cases may differ from the identity cases mentioned above in slight ways. While it is not stated outright in MacEachern (1999), it is suggested that stops and affricates at the same place of articulation in these languages can still interact in terms of ejectivity (i.e. that It, t/ can interact with Its, ts’I, etc.). Further, in Bolivian Aymara, the cooccurrence of non-identical aspirated stops is allowed, e.g. phuthu ‘hole, hollow’, t$heqha‘wing’ (MacEachern 1999:46), and in Hausa, while non- identical glottalic stops are not allowed to cooccur, other laryngeal combinations are acceptable, including heterorganic pairs of aspirated stops: keetShè ‘is split, torn’, khutthu

‘round, gourd’ (MacEachern 1999:56). While the patterns so far are not exactly what is found in Gitksan, there are a few cases which exhibit categorical patterns that mirror the gradient effects in the language. The patterns found in Gitksan closely resemble laryngeal harmonies found in some Chadic languages. For example, in Ngizim (Shuh 1978, 1997, Hansson 2001), nonimplosive stops must agree in terms of voicing. This results in a morpheme structure constraint holding over roots, and it is strictly directional, such that DVT is an allowed sequence, but *TVD is not. It should also be noted that this is an historical process where exceptions to the pattern now exist, but it is assumed that the pattern is still statistically valid.

26 It should be noted that in many such cases, the result is completely indistinguishable from a total-identity exemption that is overlaid on a dissimilatory effect (similar to what is seen in Chol).

84 (9) Ngizim laryngeal harmony kütr ‘tail’ +pü ‘clap’ tàsáu ‘find’ stti ‘sharpen to point’ gâazá ‘chicken’ dbâ ‘woven tray’ zàbIji ‘clear field’ zdü ‘six’

There is also a stronger parallel found in the Chadic family: Walker (2000a) and Hansson (2001: 150-151) present a case of laryngeal agreement in Kera whereby consonants that happen to agree in stricture also must agree in laryngeal features.

(10) Kera laryngeal harmony (data from Ebert 1979) k-màan ‘woman’ ko-taatá-w ‘cooking pot (plur.)’ ko-kámná-w ‘chief (plur.)’ go-dàar ‘friend’ g-dajgá-w ‘jug (plur.)’

Walker (2000a) and Hansson (2001:151) note that the laryngeal harmony in Kera is dependent on stricture: both the trigger and the target consonants must be plosives in order for agreement to take effect. Hansson notes that fricatives and plosives do not interact. This is entirely parallel to the Gitksan case, with some slight differences. First, the Gitksan laryngeal agreement is gradient, while the agreement in Kera is categorical. Second, the Kera case results in active alternations which affect prefix and suffix consonants; the Gitksan pattern is static, and holds over roots. Finally, the relevant laryngeal feature in the Kera case is [voice], while [cg] is the relevant feature in Gitksan. These differences aside, the same general principles are at play: there is agreement between consonants in terms of a laryngeal feature, but only if those consonants are identical in terms of stricture. Chaha exhibits a similar pattern. As Rose & Walker (2004) explain, coronal and velar stops must match for laryngeal features within roots, such that stops will be either voiceless, voiced, or ejectives.

85 (11) Chaha (Rose & Walker 2004:475) ji-ktf ‘he hashes (meat)’ ji-kft ‘he opens’ ji-dg(i)s ‘he gives a feast’ ji-drg ‘he hits, fights’ ji-t’k’ir ‘he hides’ ji-t’k’ ‘it is tight’

Rose & Walker (2004:480) also note that fricatives do not participate in this laryngeal agreement. This results in roots which can have agreeing stops (as in 11), but disagreeing

obstruents, such as /k’zw/ ‘have dysentery’ (Rose 2007), /sdI3/ ‘curse’ and /kzf3/ ‘become

inferior’ (Rose & Walker 2004). A difference here is that in Chaha the fricatives are contrastive with respect to [±voice], but in Gitksan the fricatives lack that contrast. The overall pattern mirrors that of Gitksan, though, in that the plosives, but not the fricatives, are subject to laryngeal agreement. Finally, Rose & King (2007:461) note a case of gradient laryngeal agreement in Amharic, whereby occurrences of nonhomorganic voiceless stops in “adjacent” positions that disagree with respect to [cg] are underrepresented (OlE = 0.33 for 2Ci-C OlE = 0.25 for -2C 3).C This can be contrasted with plosives that disagree in [voice] (OlE, = 1.57 for 12C-C OlE = 0.82 for )23C-C and plosives that disagree in [voice] and [cg] (O/E = 1.09 for 12C-C, OlE = 0.94 for 23).C-C This type of gradient agreement is entirely parallel to the gradient, laryngeal agreement found in Gitksan. It can be emphasized again that in the cases discussed above (Ngizim, Kera, Chaha), there is not a homorganicity requirement that is imposed on a laryngeal agreement, and it is in this respect that these languages are similar to Gitksan. The Amharic case is even more strikingly similar to Gitksan, as this is a gradient laryngeal agreement imposed on plosives, with the difference being one of magnitude: the effect is roughly 2.5 times as great in Abmaric as in Gitksan.

86 4.2.1. Analysis

The specific theory that is adopted is that of “Agreement by Correspondence” (Walker 2000ab, Hansson 2001, Rose & Walker 2004), and the idea is that two segments in an output string can be set in correspondence with each other (by means of similarity), and that correspondence constraints can force these segments to agree for a particular 27feature Agreement by Correspondence is achieved through two separate correspondence constraints:. one that demands a correspondence relation to be established between two output segments, and one that demands output correspondents agree with respect to some feature. Agreement by Correspondence is outlined schematically in (12).

(12) Consonantal correspondence model (adapted from Rose & Walker 2004:492)

Input /k’at! JO Correspondence Output [k’at] CC Correspondence

Rather than correspondence between input and output segments, Agreement by Correspondence views correspondence as also holding between output segments. CC correspondence differs from Input-Output (10) or Base-Reduplicant (BR) correspondence in that the latter two relations are given by the architecture of the grammar, whereas the former is “created” by the demands of a highly ranked or weighted C0RR-CCconstraint (which is, of course, violable). The constraint schema for correspondence proposed by Rose & Walker is as follows:

(13) C0RR-C+->C:Let S be an output string of segments and let ,1C C be segments that share a specified set of features F. If ,1C C e 5, then 1C is in a relation with C; that is, 1C and C are correspondents of one another. (Rose & Walker 2004:49 1)

This constraint demands that two segments in an output string are correspondents. In order to account for different cut-off points for agreement in different languages (i.e. the fact that

27The idea that Agreement by Correspondence constraints can be used to capture gradient patterns is also expressed in Pater (2007).

87 agreement is enforced between different consonant types) the similarity-based hierarchy of correspondence proposed by Rose & Walker (2004) is adopted. The notion of similarity that this hierarchy is based on is from Frisch et al. (2004), whereby the more similar two segments are, the more likely they will interact (in other words, the more likely they will be in a correspondence relation). C0RR-CC constraints are arranged in a type of (partial) implicational hierarchy. Correspondence between segments that differ in a number of different features, such as [son, cont, cons] implies correspondence between segments that differ in a subset of these features, such as [son, cont] or [son, cons]. Likewise, correspondence between segments that differ in these subsets of features implies correspondence between segments that differ with respect to only a single one of these features, such as [son], [cont], or [cons]. While Rose & Walker (2004:498) discuss a similarity-based hierarchy for only the range of plosives with respect to [cgJ agreement, this is not adequate to account for all of the data presented above, nor will it be adequate to account for the uvularity agreement phenomenon to be discussed in chapter 5. In order to capture the behaviors of plosives and fricatives vs. sonorants, the similarity hierarchy below must be adopted (where TIK = plosives, SIX = fricatives, NIL = sonorants):

(14) Similarity-based correspondence hierarchy CoRR-T-*K, C0RR-S->X>> CoRR-T-*X>> C0RR-N-*L >> Coiu.-T-*L

(14) constitutes an aspect of the phonology of Gitksan (though it remains to be seen wether it also reflects a fixed, universal ranking). Under this view, correspondence between plosives and correspondence between fricatives is highest ranking, followed by correspondence between obstruents. This is followed by correspondence between sonorants, and at the top of the hierarchy, the constraint demanding correspondence between any two consonants. Importantly, the labels used for each of these points on the hierarchy are meant to reflect the fact that place of articulation is irrelevant; this is especially important for the analysis of glottalization agreement, as there is no place of articulation effect that underlies the

88 agreement between glottalized segments. (14) illustrates a ranked hierarchy of constraints; these constraints are formally defined in (15-19) .28below (15) C0RR-T÷-*K:Let S be an output string of segments and let C,, C,,be segments that share the specified set of features [-son, -cont, +cons]. If C,, C’,,e 5, then C, is in a relation with C; that is, C, and C, are correspondents of one another.

(16) Colu-S4-*X: Let S be an output string of segments and let C,, C be segments that share the specified set of features [-son, +cont, +consj. If C, C,, e S. then C, is in a relation with C; that is, C, and C, are correspondents of one another.

(17) C0RR-T÷->X:Let S be an output string of segments and let C,, C be segments that share the specified set of features [-son, +consl. If C,, C e S. then Cx is in a relation with C; that is, C, and C, are correspondents of one another.

(18) CoRR-N->L: Let S be an output string of segments and let C,, C’,,be segments that share the specified set of features [+son, +cons]. If C, C, e 5, then C, is in a relation with C; that is, C and C, are correspondents of one another.

(19) CoRR-T->L: Let S be an output string of segments and let C,, C, be segments that share the specified feature [+cons]. If C, C e S, then C is in a relation with C; that is, C and C,, are correspondents of one another.

Most of these C0RR-CC constraints are in a subset relation. For example, CORR T÷-*Xapplies to obstruent-obstruent pairs, regardless of the values for [±cont]. Thus, this particular constraint will apply to T-K and S-X pairs as well as T-X pairs. In similar fashion, a constraint like C0RR-T÷-,Lapplies to all consonant pairs, regardless of values for [±cont] and [±son]. This would include T-K, S-X, N-L, T-X, S-L, T-L, etc. The point here is that each CoiR-CC constraint defines a similarity threshold, where correspondence is required of anything above that threshold. This results in a type of stringency relation (cf. de Lacy 2002, 2004), however, C0RR-T÷-*X>> CoRR-N->L is separate from this issue, and is a stipulation. It could be the case that languages will conflate this ranking, such that correspondence is

28 The feature [+cons] is used in these constraint defintions not to exclude any particular class of segments, but rather exists to provide a target for the constraint in (19) to apply to. Since [-son] consonants are also redundantly [+cons], the feature is used for consistency in all CoRR-C-*C constraints. This also illustrates the point that (15)-(18) all stand in a subset relation with (19).

89 established equally between obstruents, or between sonorants. The Gitksan data presented in chapter 2, however, motivates this stipulative ranking based on the fact that the obstruents exhibit agreement effects, whereas sonorants do not. Since the only agreement effects that are encountered are those between 29obstruents in the language (see section 2.2.2.2 of chapter 2 and section 5.2 of chapter 5), the relevant cutoff point in Gitksan is below CoRR-T->X. The effects of all of these constraints are displayed in (20). For each pair of consonants that are at the same threshold of the hierarchy in (14), when those consonants are not in a correspondence relation (indicated by subscript letters), a violation is incurred. Compare the (a) vs. (b) candidates (where CoRR-T÷L is shown to illustrate that all consonants are set in correspondence):

(20) Placing consonants in agreement — violation profile CoRR-T-*K CORR-T->X C0RR-T4->L 1 a. t-k lbt...t’-.x_y * * * 2 a. 2b.t-x * * 3 a.

1 * .J2bf . 1yLx_

Since glottalization among the sonorants does not trigger an agreement effect, and since fricatives are not specified for [cg], the most relevant of these constraints is that responsible for placing obstruents into correspondence (T->X). This also means that the highest CORR-CCconstraint that N-L pairs would fall under (the relevant constraint is C0RR- N*+L) would have to be below the threshold defined by the constraint ranking. Within OT, a highly ranking CoFu-T->X constraint requires all output obstruents to be in correspondence. Under the Agreement By Correspondence view, segments that are required to be in correspondence in (14) are then further required to agree in their values for specific features. In this particular case, the relevant feature is [cg]. There is a set of correspondence

29 In actuality, fricatives are not specified for [cg], so do not participate in the glottalization agreement. This is forreasonsindependent of the agreement effect. For this particular analysis, it would suffice to consider the “cutoff’ point to be below CORR-T4-*K,which would demand that only plosives be in agreement. However, anticipating the analysis of the uvularity agreement in chapter 5 where all obstruents participate in the pattern, CoRR-T->X will be the cutoff point used here, in the very least for consistency between the two analyses.

90 constraints that ensures agreement among correspondents for a particular feature. The general formulation of the correspondence constraint is as follows:

(21) IDENT-CC[F]:Let C, be a segment in the output and C,, be any correspondent of C, in the output. If C is [aF] then C, is [aF].

Since it has been established that there is a gradient glottalization agreement in Gitksan, the relevant feature is [constricted glottis]. Thus, the constraint schema in (21) can be filled in appropriately with the feature [cg]:

(22) IDENT-CC[cg]:Let C, be a segment in the output and C, be any correspondent of C, in the output. If C is [cg], then C, is [cg]. (Rose & Walker 2004:499)

The demands of IDENT-CC[cg] will confLict in some cases with typical Input-Output correspondence constraints. The relevant constraint here is IDENT-lO,which will demand featural identity between input and output correspondents:

(23) IDENT-IO[cg]:Let C, be a segment in the input and Cf,,be any correspondent segment of C in the output. If C is [cg], then C is [cg].

Since this analysis is assuming [cg] to be a privative feature, it is necessary to have an Output-Input version of the constraint, as well (cf. Pater 1999):

(24) IDENT-OI[cg]:Let a be a segment in the output and 13be any correspondent segment

of a in the input. If a is [cg], then 13is [cg].

IDENT-IO[cg]prevents deglottalization, and IDENT-OI[cg]prevents glottalization of a plain consonant. These constraints are an important component to the present analysis, as they will be the constraints that show the effects of maximal weighting in the Harmonic Grammar approach developed in chapter 3. For expository purposes, an OT approach with strict domination will be entertained for the moment.

91 The ranking necessary for agreement is C0RR-C+->C,IDENT-CC[FJ>> IDENT-IO[F]. This is illustrated in the tableau in (25) (with hypothetical forms, and with similar subscripts indicating a correspondence relation in output forms):

(25) Ranking for agreement Ik’atl CORR-T->X IDENT-CC[cg] IDENT-IO/OI[cg] ja. ‘k’at’ * b. k’at c. k’at *!

Since the output correspondence constraints C0RR-T4->X and IDENT-CC[cg] are ranked highest, the output candidate must satisf’ these constraints by having its obstruents in correspondence, and also have those correspondents agree with respect to [cgj. Candidate (a) satisfies these constraints at the expense of lower-ranked IDENT-OI[cg],while candidates (b) and (c) both violate either CoRR-T÷-.X by virtue of not being in a correspondence relation, or IDENT-CC[cgj by not having obstruents that agree for the feature [cgj. This is the only ranking that enforces agreement; this means that non-agreeing inputs (like /kat’/ in 25) will be made to agree in the output ({k’at’] in 25). Ranking IDENT-IO[F}above IDENT-CC[FJ and C0RR-T+->Xwould not derive an agreement effect; rather, surface forms faithful to the input would be selected every time. This is best illustrated with a mixed ejective-pulmonic pair as an input form:

(26) Agreement blocked /k’at/ - IDENT-IO[cg] IDENT-OI[cgj IDENT-CC[cg] CORR-T+->X a. k’at’ *! b. k’at’ *! * c. k’at * ‘d. k’at *

Candidates (a) and (b) are eliminated due to their violations of highly ranking IDENT-OI[cg]. This leaves candidates (c) and (d) as the winners here, although this shouldn’t be taken literally to mean that these candidates tie. Rather, which candidate wins will depend on how IDENT-CC and C0RR-T+÷X are ranked with respect to each other, and this cannot be

92 determined at the moment, due to the two candidates being phonetically identical. This issue will be taken up again shortly in the context of the learning algorithm. In order to trigger harmony in those forms that exhibit it, IDENT-CC[cg]must outrank the relevant faithfulness constraint responsible for maintaining identity between input and output (IDENT-lO/IDENT-Ol)(that is, still assuming a highly ranked CoRR-T->X). In this way, whether or not the input form is an ejective-ejective (27) or mixed pulmonic-ejective pair (28), the result will be the same.

(27) Glottalization agreement (underlying agreeing) /t’ik’/ CoRR-T-X IDENT-CC(cg) IDENT-IO[cg] IDENT-OI[cgJ

+‘ •1 *, * L

(28) Glottalization agreement (underlying disagreeing) It’ik/ C0RR-T+-*X IDENT-CC(cg) IDENT-IO[cg] IDENT-OI[cg] c’j’ * +, •1 *, L x”x

Putting hypothetical examples aside, it is now time to explore the actual numerical weightings that are assigned to each of these constraints. Up to this point, it has been assumed that the phonological phenomenon has been a categorical one (such as it would be in Chaha); this however, is not true. As noted above, the pattern of laryngeal agreement in Gitksan is gradient, but it is also important to reiterate the fact that it is static as well (i.e. it does not result in alternations). This being the case, output forms are taken to be identical to input forms, by lexicon optimization. Granting a higher numerical weighting to the faithfulness constraints governing input-output relations (IDENT-IO[cgjand IDENT-OI[cgj) ensures that output forms match input forms. This is because this model is not meant to capture an active pattern where there are alternations, and it does not posit underlying forms that are not identical to output forms (by lexicon optimization). This is generally true for both HG and phonotactic learning in OT (Hayes 2004, Prince & Tesar 2004). Instead, following the example set in chapter 3, output forms are identical to input forms, and (nearly) all markedness constraints are more lowly weighted and play the role of tracking tendencies over the lexicon. The markedness constraints incur violations, and these violations

93 contribute to lower harmony scores for candidates that violate them. IDENT-IO[cgJand IDENT-OI[cgj could in theory afford to be lowly weighted for the following reason: in principle, a relatively similar pair of segments could violate a gang of C0RR-CC constraints and beat the relevant IDENT-lOand IDENT-Olconstraint. Despite this, it is always possible to satisfy those constraints by imposing correspondence “invisibly”; that is, violating the lowly- weighted IDENT-CCinstead of the entire gang of C0RR-CC constraints. An interesting problem arises here: how does a learner distinguish between an output that does have a correspondence relation but violates agreement vs. an output that does not have a correspondence relation but vacuously satisfies the constraints on agreement? These two cases would be phonetically identical, but the different learning paths that they would trigger would be wildly different. Of course, there is no evidence that plosives are necessarily in a correspondence relation, as agreement is not imposed on all pairs of plosives. This is potentially where the role of analytic bias (Moreton 2008) would select one structure over another, or some other heuristic would guide the learner, such as choosing the learning path that involves adjusting the fewest constraint weights. The idea is that the learner comes pre equipped with a preference for non-agreeing structures to be in non-correspondence, rather than in non-agreeing correspondence. This contrast is illustrated in (29), where phonetically indistinguishable output candidates violate different constraints.

(29) Possible structures for non-agreeing outputs Input: k’-t CoRR-T÷÷X IDENT-IO[cgj 1%xLy1’ + * * k’XLX

There is a problem inherent in selecting corresponding output structures (and hence, with a highly weighted CORR-CCconstraint in the hierarchy): As will be discussed below, all C0RR-CC constraints start with the same weight in the initial state. If correspondence is always respected by means of satisfying these C0RR-CC constraints, then they will not be demoted in weight (or promoted in weight, for that matter); they will all remain at the same weight until convergence. Thus, C0RR-N&*L,which penalizes non-correspondence between sonorants will weigh the same as CoRR-T-+X, which penalizes non-correspondence between sonorants. But if this is the case, then the generalization that T’-T is underrepresented (and

94 penalized, and less acceptable), while N-L’ is not cannot be captured by the grammar because these constraints are making the exact same contribution to harmony scores (and by extension, acceptability scores).

(30) Disagreeing sonorants Weight 100 100 50 Input: n-i’ CoRR-N<-L CORR-T-*X IDENT-CC[cg] H A “n-l’ -1 -50 50 n-l’ -i -100 -50

(31) Disagreeing obstruents Weight 100 100 50 Input: t-k’ CORR-N-*L C0RR-T÷*X IDENT-CC[cg] H A t-k’ -1 -50 50 t-k’ -1 -100 -50

In the examples above, disagreeing obstruents and sonorants receive the exact same harmony scores, and the exact same acceptability scores. In this way the constraint weights have failed to capture the difference in how the two consonant types pattern in the lexicon: disagreeing obstruents are underrepresented, and disagreeing sonorants occur at chance. This necessitates the need for the C0RR-CC constraints to be nudged downward during learning, while IDENT-CCwould stay at its original weight. There is also a typologically interesting note to be made here: if IDENT-CC[cgj were effectively “undominated” (i.e. no lower-weighted constraints could gang up to overpower it), then agreement would be imposed on T’VK sequences (for which all three of the relevant Corr CC constraints could gang up on IDENT-OI[cgj),but agreement would not be imposed on T’VL, as the relevant correspondence constraint (CoRR-T-*L) would be too lowly weighted to have any effect. The assumption that agreement is highly weighted, though correspondence is less strongly imposed (and hence agreeing structures don’t always surface) follows similar approaches by Hansson (2001), Rose & Walker (2004) and McCarthy (2007). Just as with the discussion of the learning algorithm in Chapter 2, learning the gradient laryngeal agreement will proceed here step-by-step, starting with the initial state. Again, assuming a weight for the faithfulness constraints (IDENT-IO[[cg]and IDENT-OI[cgj)

95 of 50, and a weight of 100 for all markedness 30constraints (CORR-T*-*K,C0RR-N+->L, CoRR-T-*X, CoRR-T.->L,IDENT-CC[cg])and a plasticity of 1 (where plasticity is simply the

amount by which constraints are reweighted), learning proceeds as follows. For Stage 1, where constraints are weighted as in the initial state, a hypothetical case of a T’-T input is given to the learner (now with real Gitksan roots as examples):

(32) Datum 1: T’-T input Weight 100 100—> 100—> 100—> —50 ‘c—50 H /k’ats/ ‘land, IDENT CORR CORR CORR IDENT IDENT arrive’ CC[cg] T->K T->L IO[cg] OI[cgj k’ats -1 -100 ‘icats 1 50 k’ats’ -1 -50 k’ats -1 400 Vk’ats -1 -1 -1 -300

Since, as noted above, there is a bias towards selecting forms that do not agree based on their non-correspondence, then such a datum would trigger the re-weighting (in a downward direction) of all C0RR-CC constraints. IDENT-CC[cg] plays no role in the selection of either the correct output or the error(s), so it remains at the same weight. And since the errors violate IDENT-IO[cg] and IDENT-OI[cg](by their satisfaction of IDENT-CC[cg])and the correct output does not, these constraints are nudged up in weight. Learning can then proceed with another hypothesized input, one with another disagreeing pair of consonants:

30As McCarthy (2007) notes, since CC-constraints and1DENT-CCconstraints evaluate the well-formedness of output structures, they are technically markedness constraints.

96 (33) Datum 2: T’-T input Weight 100 99—> 99—> 99—> +—51 *—51 H /boq’/ IDENT CORR CORR CORR IDENT IDENT ‘be lame’ CC[cgj T->K T-*X T÷->L IO[cg] OI[cg] boq’ -1 -100 boq -1 -51 ‘p’oq’ -1 -51 Vboq’ -1 -1 -1 -297

Much the same result obtains here, as the same constraints are affected in the same directions. So far the weights for the C0RR-CC constraints have been nudged down with each datum, while IDENT-IO[cg] and IDENT-OI[cg] have been nudged up in weight. So far the C0RR-CC constraints have behaved uniformly. However, an input-ouput pairing of T-L would violate only C0RR-T4->L,thus nudging down the weight for this constraint. This is illustrated with a T’-L form, just to make the point that agreement in values for [cgj is irrelevant:

(34) Datum 3: T’-L input Weight 100 98 98 98—> —52 ÷—52 H /m’ats/ IDENT CORR CORR C0RR- IDENT IDENT ‘hit, CC[cgj T(—>K T->X T->L IO[cg] OI[cg] strike’ m’ats -1 -100 ‘m’ats’ -1 -52 mats -1 -52 Vm’ats -1 -98 m’ats’ -1 -1 -150 mats -1 -1 -150

The table in (35) lists the constraint weights at the initial state, and after the first stages of learning:

97 (35) Constraint weights after each stage of learning Constraint Initial State Datum 1 Datum 2 Datum 3 IDENT-CC[cg] 100 100 100 100 CORR-TK 100 99 98 98 CORR-TX 100 99 98 98 CORR-TL 100 99 98 97 IDENT-IO[cg] 50 51 52 53 IDENT-OI[cg] 50 51 52 53

Of course, there are other possible consonant pairs that would be encountered by the learner. These include pairs that fully agree for [cg], as well as pairs that have no specifications for [cgj (and hence also agree for [cg]). An ejective-ejective pair is given in (36), and a pulmonic-pulmonic pair is given in (37). Given that the candidate with the highest harmony score for each case is the expected output, there is no error, and thus there is no learning triggered. Since there is no learning, the constraint weights do not change.

(36) Datum 4: T’-T’ Weight 100 98 98 97 53 53 H It’ak’I IDENT CORR CORR CORR IDENT IDENT ‘fold’ CC[cg] T-K T-X T-*L IO[cgj OI[cg] c’Vt’ak’ 0 tak -2 -106 t’ak -1 -l -153 t’ -1 -1 -293 xak’y

(37) Datum 5: T-T input Weight 100 98 98 97 53 53 H /bft/ IDENT- CORR- C0RR- C0RR- IDENT- IDENT ‘divorce’ CC[cg] T+->K T(->X T->L IO[cgj OI[cg] crvbft 0 ‘it -1 -1 -153 ‘it’ -2 -106 bit’ -1 -1 -153 bit -1 -1 -1 -293

The learning algorithm continues in this fashion until convergence, where the weights for each constraint will be fixed, and learning will cease. The learning algorithm was

98 implemented using Praat (Boersma & Weenink 2007) as in Chapter 2, with the modification that noise in evaluation was set to 0.0 (due to some of the constraints resulting in weights that are close enough to each other that noise in the data could result in a weight reversal). The fmal weights of the constraints are listed in (38):

(38) Constraint weights after learning Constraint Weight after learning IDENT-OI[cg] 100.5 IDENT-IO[cg] 100.2 IDENT-CC[cgj 100 C0RR-S+-*X 100 CORR-N->L 80.6 CORR-T-*K 60.8 CoRi-T-*X 37.3 CORR-T-*L 0.3

The implementation of the algorithm involved the gradual nudging down of the markedness constraints, and a gradual nudging up of the faithfulness constraints until nearly all markedness constraints (aside from IDENT-CC[cgj)are weighted lower than faithfulness (and for some constraints, low enough to prevent any ganging up with other constraints to overtake a faithfulness constraint). Importantly, just as with the OCP effects discussed in chapter 3, IDENT-IO[cg]and IDENT-OI[cg]are crucial to the analysis. The gradual increase in the weight of these constraints allows for lexical items to surface identical to their underlying forms, while the weight of the markedness constraints essentially tracks the violations that are incurred by these lexical forms (by means of the decrease in their weights). This approach allows both the lexicon and the phonological grammar (where ‘lexicon’ refers to the frequency distribution patterns in the lexicon, and ‘phonological grammar’ refers to the input-output mappings that are defined by the grammar) to be learned gradually, and crucially, it results in a state which mirrors the structure of the lexicon — in this case, with respect to the gradient laryngeal agreement. This is illustrated in (39-41), where acceptability scores are presented for all combinations of plosives.

99 (39) t-k Weight 100.5 100 60.8 37.3 0.3 Input: IDENT- IDENT- CORR- CORR- CORR- H A /t4cJ OI[cg] CC[cgj T€*K T->X T+->L t-k 0 98.4 t-k’ -1 -1 -1 -98.4 t’-k’ -1 -100.5

(40) t-k’ Weight 100.2 100.5 100 60.8 37.3 0.3 Input: IDENT- IDENT- IDENT- CORR- CORR- CORR- H A It-k’! IO[cg] OI[cgj CC[cg] T÷->K T÷->X T->L t-k -1 -100.2 t-k’ -1 -1 -1 -98.4 1.6 t’-k’ -1 -1005

(41) t’-k’ Weight 100.2 100.5 100 60.8 37.3 0.3 Input: IDENT- IDENT- IDENT- CORR- C0RR- C0RR- H A !t’-k’/ IO[cgj OI[cgj CC[cg] T-*K T÷-*X T->L t-k -2 -201 t-k’ -1 -1 -1 -1 -198.6 t’-k’ 0 198.6

The acceptability scores of these pairs of consonants can be compared, along with their respective O/E values. There is a correlation between Acceptability and attestedness in the lexicon: the higher the Acceptability, the higher the O/E.

(42) Comparing acceptability for plosive pairs Consonant pair Acceptability O/E T-K’ 1.6 0.75 T-K 98.4 1.13 T’-K’ 198.6 1.75

The same comparison can be made with disagreeing sonorants:

100 (43) Disagreeing sonorants Weight - 100.2 100.5 100 80.6 0.3 Input: IDENT- IDENT- IDENT- C0RR- CORR- H A In-i’! IO[cg] OI[cg] CC[cg] N+-L T-*L n-i -1 -100.2 n-l’ -1 -1 -80.9 19.3 n’-l’ -1 100.5

These values can now be placed in the context regarding the statement above concerning the grammar capturing the differences between the various consonant types: the Acceptability scores for disagreeing plosives is lower than for disagreeing sonorants, which is a reflection of the degree of attestedness of these pairs of consonants in the lexicon, which is indicated by their OlE values:

(44) Comparing acceptability between obstruents and sonorants Consonant pair Acceptability OlE T-K’ 1.6 0.75 N-L’ 19.3 1.06

Harmonic Grammar also captures different markedness dimensions, as well. For example, [t’-tsj is more underrepresented (and will presumably be judged worse) than [k’-ts] due to the contribution of the OCP-coR constraints to the harmony/acceptability scores (cf. discussion of this in section 3.3.3 of Chapter 3). There are several desirable aspects of this analysis. First, this analysis captures the gradient glottalization agreement patterns in the same way as the OCP effects discussed in chapter 3: by means of numerically weighted constraints. Second, the actual assimilatory effects are derived by agreement by correspondence, a mechanism that has been used by various authors to account for consonant harmony effects (Hansson 2001, Rose & Walker 2004), as well as glottalization agreement effects (Hansson 2004, Rose & Walker 2004, Coon & Gallagher 2007).

Another desirable result is how the learning algorithm produces a final grammar which reflects the probabilistic agreement patterns in the lexicon. At each stage of learning, the algorithm re-weights constraints. This effectively “shapes” the grammar with each new datum. At convergence, the grammar winds up in its end “shape”, which because the algorithm uses lexical items to reweight constraints, reflects the structure of the lexicon. The

101 specific way this is manifested is through the resulting harmony scores that candidates receive (or more exactly, through the resulting acceptability scores). Fully faithful agreeing outputs will receive higher harmony/acceptability scores than those of fully faithful disagreeing outputs. This is the link to the empirical observation that agreeing items are overrepresented in the Gitksan lexicon, while disagreeing items are underrepresented. Thus, the learning trajectory boils down to learning the lexicon piece-by-piece. For a language that shows no under- or overrepresentation for disagreeing/agreeing pairs of glottalized consonants, the learning procedure would be exactly the same, but the resulting constraint weights would be different to the point that harmony scores (and hence, acceptability scores) would not reflect any probabilistic lexical agreement or disagreement, and both agreeing and disagreeing pairs would surface at 50% probability. This result would obtain if the C0RR-CC constraints were nudged far enough down in weight (triggered by the presence of non- agreeing forms) that the impact that these constraint weights have on harmony and acceptability scores would be negligible, which is a contrast from the Gitksan-type of system. Next, some categorical restrictions on cooccurring laryngeal features within the domain of the segment will be discussed.

4.3. Paradigmatic Restrictions on Laryngeal Features

This section investigates some intrasegmental, or paradigmatic restrictions on the occurrence of laryngeal features within obstruents. While the previous section dealt with gradient lexical patterns, this section deals with categorical restrictions. Despite this, the same basic framework can be employed: Harmonic Grammar. Another reason for discussing these paradigmatic cooccurrence restrictions is to draw attention to the role that gradient phonotactics can play in informing the analysis: while there is little overt “phonological” evidence available to show that voicing is an active phonological phenomenon (rather than just phonetic), cooccurrence restrictions can help to classify how voiced plosives behave with respect to other plosive types. First the phenomenon of plosive voicing will be established, and the phonological reality of [voice] will be discussed. The constraints necessary for voicing will be introduced. This will be followed up with a discussion of the other laryngeal features: [cg] and [sg]. It

102 will be noted that in each case there are restrictions on what laryngeal features can be realized on obstruents, and that the blocking of obstruent voicing is due to specifications of other laryngeal features.

4.3.1 Obstruent voicing

The focus of this section is on the well-documented process of obstruent voicing which was mentioned earlier in section 4.1. This alternation is discussed because it is central in demonstrating the paradigmatic restrictions on laryngeal features. This pre-vocalic voicing affects stops and affricates, but fails to affect fricatives, ejectives, and a small class of exceptionally aspirated stops. The possessive paradigm that is famously given to illustrate the dynamics of obstruent voicing in Gitksan is presented below (data from Rigsby 1986:133):

(45) Obstruent voicing paradigm a. mo-pip! ’3 [nI.bIp] ‘maternal uncle’ b. !no-pip-oj‘I [rn.bi.biLl ‘my maternal uncle’ c. mo-pip-on [ni.bI.binj]1 ‘your (sg.) maternal uncle’ d. !no-pip-tJ [nI.bipt] ‘his, her maternal uncle’ e. /no-pip-om’/t [m.bI.bI?m] ‘our maternal uncle’ f. !no-pip-som’I [ni.bip.si?m] ‘your (p1.)maternal uncle’ g. /no-pip-ti:t/ [nI.bIp.di:t] ‘their maternal uncle’

From the above data, it is clear that plosives are voiceless in pre-obstruent (45d,g) and in final (45a,d,g) positions. Voiced plosives appear in pre-vocalic position (the relevant alternations of the final plosive in nibip are found in 45b,c,e). Furthermore, the direction of voicing is always regressive. Cases of plosives to the right of vowels show that the voicing process is not progressive:

31 is listed in Rigsby (1986) as simply a prefix. Ipip! is the root for maternal uncle, and the suffixes mark person and number agreement.

103 (46) Post-vocalic plosives /?aks/ [Paks] ‘water’ /tap-t/ [dapt] ‘his, her liver’ /matx/ [matx] ‘mountain goat’ hit! [lIt] ‘wedge’

This set of alternations in voicing has led Rigsby (1968, 1986), Hoard (1978), and Rigsby & Ingram (1990) to posit voiceless plosives in the consonantal inventory of Gitksan, with a voicing rule to derive the voiced allophones. In contexts with plosive clusters, it is the plosive directly preceding the vowel which becomes voiced. If an obstruent isn’t strictly adjacent to (and preceding) a vowel, it does not become voiced. This is most obvious in medial plosive clusters:

(47) Medial plosive clusters !sqapti:+! [sGapdi:1] ‘among’ !t’ak-ti:t/ [t’akdi:t9 ‘they forgot it’ /lit—ti:t/ [litdi:t9 ‘their wedge’

These examples raise the possibility that it is only onsets which undergo the process; however, the process does not affect members of onset clusters which are not adjacent to a vowel (48).

(48) Word-initial plosive clusters !ptal/ [pdal] [pdl] ‘rib’ /ptu’?/ [pdo?°] ‘door’ /pte:q/ [pde:q] ‘clan’ ‘tribe or phratry’ /pto:4i [pdo:ij ‘siding’

This indicates that if it is a condition on onsets, there are additional restrictions involved which force only prevocalic members of a complex onset to undergo voicing. This pattern of voicing runs counter to Greenberg’s generalization (1966) that if a language has C clusters that disagree in voicing (i.e. TD), then it has C clusters that agree in voicing (i.e.

104 TT). As pointed out by Blevins (to appear), there are several gross probabilities for the distribution of word-initial consonant clusters cross-linguistically, including #TT, #TD, #DT, and #DD. Blevins notes that #TD sequences should be less common than #TT (since many languages lack a voicing contrast and won’t have D at all, but more common than #DD (since voicing more difficult to maintain over these sequences). Blevins also notes that a system with only #TD can evolve from a #TT system by means of assimilation in voicing with a following vowel (pg. 7) *#‘F’I’V> #TDV. This is likely the diachronic explanation for the pattern of plosive voicing found in 32Gitksan though it is unknown to me if there are any other languages that have undergone this, type of diachronic change. Finally, there is some question as to whether plosive voicing is triggered by a prevocalic or a pre-sonorant context. As it happens, there are no plosive-sonorant sequences available in the 33language (Brown to appear), aside from those derived by reduplication (49).

(49) Examples of derived plosive-sonorant sequences iaX-iip-ii ‘to roll (intrans) (p1)’ lip-lo?op1p ‘rock, stone (p1)’ mit-mitx”’ ‘to be full (p1)’

There are two observations to be made about the data above: first, there is no voicing of a plosive followed by a sonorant consonant. Second, there are morpheme boundaries between all of the plosive plus sonorant sequences in these examples. Given that these are reduplicants, it seems reasonable to suggest that the plosives are subject to Base-Reduplicant correspondence, and that a constraint such as IDENT-BR[voice]prevents voicing from affecting consonants in the reduplicant. However, since deglottalization of an ejective in the reduplicant results in a voiced plosive (dicussed at length in chapter 6), then BR- correspondence cannot be responsible for blocking plosive voicing in these contexts (as

32 To my knowledge, both Nisgha and Cost Tsimshian both exhibit the same prevocalic voicing of plosives. There are two genuine exceptions to this, [libleet] ‘priest’ and [gloq] ‘shame’. The first is a loanword, the second is most likely involves a reinterpretation of [dl’] into a [gil sequence. There are also several prima fade exceptions involving a labialized velar [gWn followed by the coronal nasal (i.e. . .1. These have been shown in

Brown (to appear) to be reduced forms of [g”’in. .1. Many orthographic forms (as found in the appendix) have ; this indicates a fronted velar, and not a true [gj] sequence.

105 voicing in the base, as in the deglottalization case, would be a violaton of this constraint). Thus, it seems reasonable to assume at this point that plosive voicing is not triggered by sonorant consonants. With these facts in hand, it is necessary to define the markedness constraint that forces plosive voicing. A constraint like that in (50) can be formulated, which is a positional paradigmatic constraint on feature cooccurrence (Prince & Smolensky 1993, Pulleyblank 1997):

(50) OBsVoIcE: prevocalic [-son] segments are [+voice]

This constraint must dominate a faithfulness constraint in order to trigger this process in the language; presumably this constraint is simply IDENT-OI[voice],which bans changes from input to output with respect to the feature [voice]. Since the privative feature [voice] is being assumed here, then prevocalic plosives acquire a Laryngeal node in the output, but non prevocalic plosives stay bare with respect to this node (see examples 5a and 6 above). It is tempting to derive this type of assimilation in the same way as the gradient laryngeal agreement, that is, by the agreement by correspondence paradigm. However, this makes the wrong predictions for gradient laryngeal agreement in that correspondents would have to include all consonants and vowels. Concretely, this means that sequences such as [...pd...] would not be allowed. Since [p1 and [d] are by definition more similar than [d] is to any vowel, and since [d] is categorically forced to be voiced in prevocalic position, then [p] would also be voiced preceding [d]. This is not the case, as plosives in pre-consonantal position are not voiced, and words like [pdal] ‘rib’ exist in the language. This problem is discussed further in section 5.3.1 of chapter 5. In order to determine whether voicing is phonologically active in the language, rather than merely a phonetic effect, consonant cooccurrence numbers can be observed. Since a great many roots are 1C 2VC in structure, these sequences are ideal for testing whether the role of [voice] is active, assuming that “pulmonic” in Cl = [D] whereas “pulmonic” in C2 = [T] -- which is going to be true in the vast majority of the CVC sequences in the data. It can be seen that voiced-voiceless are about what would be expected under chance: O/E = 0.87 (n.s.). So far, this tells us little about the status of voicing in the language. However, when

106 cooccurrence with ejectives is taken into consideration, the picture changes considerably. D T’ sequences are very highly underrepresented, with OlE 0.28 (<.001), while T’-T sequences show no discernible effect (OlE = 1.21, n.s.). The cooccurrence numbers for all plosive types in CVC sequences, including ejective-ejective (T’-T’) are given in (51).

(51) Cooccurrence of plosive types in all CVC 34sequences D-T’ 0.28*** (4/14.3) D-T 0.87 n.s. (56/64.3) T’-T’ 2.41** (15/6.2) T’-T 1.21 n.s. (34/28)

What these results indicate is that in CVC sequences, pairs of consonants which disagree in both glottalization and voicing (DVT’) are severely penalized, as they are by far the most underrepresented. At the other extreme, consonant pairs which agree for both of these features (T’VT’) are strongly favored. Pairs that agree for one feature but disagree for the other (DVT, T’VT) are intermediate between the two extremes, with no significant effect. These results are interesting, for the following reasons. The overall rarity of DVT’ sequences suggests that there exists a gradient constraint against cooccurring plosives that have conflicting laryngeal feature 35specifications This approach assumes that both [voice] and [cg] are privative features, and that T is unspecified for laryngeal features. This is not entirely an unexpected state of affairs. There are several cases of cooccurrence restrictions based on [spg] and [cg]. For instance,. MacEachern notes that both Cuzco Quechua and Peruvian Aymara prohibit aspirated stops and ejective stops from cooccurring (i.e. *tapha, 36*t’atlla). There are also cooccurrence restrictions on [voice] and [cg], such as in Chaha. Historically in Chaha, DVT’ > T’VT’, giving rise to an agreement pattern (Rose & Walker 2004; see also the discussion of Chaha earlier in section 4.2). The main reason why the results are interesting, however, has to do with the phonological status of voicing in Gitksan. Voicing in the language is allophonic, and not

The sequences DVD and T’VD are also both possible in theory, since 2C will occasionally be prevocalic (i.e. in sequences such as CV).VC Alternatively, this could12 also simply be the result of an ordering restriction, whereby C’-C is preferred over C-C’. Further investigation of the lexicon will be informative in this regard (not only in terms of laryngeal ordering constraints, but in terms of ordering constraints generally). 36 MacEachern (1999:45) notes that Bolivian Aymara exhibits a similar pattern, though with the added restriction that the stops be homorganic.

107 phonemic for the obstruents. Under an output-oriented framework, where output forms are those under consideration, it would be the surface occurrence of [voice] that would violate well-formedness constraints. Since evalution of well-formedness constraints is concerned with output representations, and since by hypothesis allophonic voicing is found only on the surface, an OT or HG analysis predicts that it would have an effect in cooccurrence restrictions. One further potential piece of evidence (not related to cooccurrence restrictions) which suggests that [voice] is phonologically active for obstruents is the maintenance of plosive voicing in reduplicants. There is a general process of deglottalization which affects reduplicants in the language (Brown 2007), whereby any glottalized consonant in the base has a deglottalized correspondent in the reduplicant. This is illustrated in (52) for the sonorants:

(52) m’asx” mis-m’asx” ‘to sting (trans.)’ m’ats mis-m’ats ‘to hit, strike’ m’axs ma:-m’axs ‘pants’ w’in win-w’in ‘tooth’

Thus, while it can be shown that glottalized consonants can become deglottalized in certain contexts, there is never a case of devoicing. Furthermore, when plosives in the reduplicant are deglottalized, they are subject to the process of plosive voicing (recall that ejectives fail to undergo prevocalic voicing):

(53) t’is dis-t’is ‘to push’ ls’ak’ cki-ts’ak’ ‘dishes (of one kind)’ kwo:tx gwitkwo:w ‘to be lost, gone’ q’ots Gas-q’ots ‘to cut’

Of course, this process could very well be evidence that voicing is in fact not active in the phonology since it appears that the reduplicative morphology can’t regulate or affect the feature. A more convincing piece of evidence for the phonological (vs. phonetic) status of

108 [voice] comes in the class of segments which fail to voice: the fricatives (to be discussed further below in section 4.3.2.3). While it has been demonstrated that affricates, which have a fricated release period, do undergo prevocalic voicing, fricatives do not. If prevocalic voicing were simply a coarticulatory phonetic effect, then it would be expected that there should be no difference between fricatives and affricates with respect to voicing. Instead, the phonology must be able to restrict the occurrence of [voice] on a particular class of features as opposed to another.

4.3.2. Blocking allophonic voicing

There are three main classes of segments that resist prevocalic voicing. These are the ejectives, the aspirated stops, and the fricatives. Each of these segment types will be described in turn.

4.3.2.1. Ejectives

We can further define the natural class of segments that undergoes the voicing process by eliminating the set of ejectives from the picture. There has been much debate in the literature as to whether ejectives undergo the voicing rule. In a much-cited paper, Hoard (1978) claimed that the ejectives in Gitksan do undergo voicing, deriving a set of implosive stops. This position has been argued against by Rigsby (1986), Rigsby & Ingram (1990), and Ingram & Rigsby (1987). While Hoard’s arguments were based on impressionistic observations, Ingram & Rigsby (1987) performed an acoustic analysis of the ejectives, with the finding that there is no voicing prior to the onset of modal voicing for the vowel. Ingram & Rigsby’s findings are in agreement with my own auditory observations of the ejectives in the language. Thus, ejectives are immune to the rule of plosive voicing discussed earlier in this section.

(54) Ejectives resist voicing si:p’in *si:6in ‘to like (a person)’ t’a:+ *cfa:+ ‘to pick (berries)’ k’ats *çfa ‘to land, arrive’

109 k’”adix* cradix ‘tojump around, flop’ q’a:t’ *.ja:t’ ‘cane’

We can now begin to answer why ejectives, while [-cont], do not undergo prevocalic voicing. It is uncontroversial that ejectives are specified with the laryngeal feature [cg], as assumed earlier in this chapter. Building on the analysis presented below for the fricatives, it can be hypothesized that since there is already a laryngeal feature present (i.e. [eg]), it serves to block the spreading of [voice]. Thus, constraints that militate against the doubling up of laryngeal features, e.g. paradigmatic constraints on feature combinations (*{spread, voice}; * {constricted, voice}) must be highly ranked (i.e. heavily weighted) in the language.

4.3.2.2. Aspirated Stops

This section discusses the role that the feature [sg] plays in Gitksan. There are two primary reasons, aside from the existence of the glottal fricative [h], for adopting [sg]: the existence of words with stops which do not undergo prevocalic voicing (and instead are aspirated), and the resistance of fricatives to voicing. It will be claimed that aspirated stops and fricatives are underlyingly specified for [sg], and that these segments do not voice due to their being [sg], as well. A constraint on multiple laryngeal features associated to a single root node blocks voicing in each case. This constraint is related to the one mentioned in the previous section on cooccurrence restrictions. There are noted exceptions to the process of obstruent voicing. For instance, Rigsby notes that many loans from English like [ti:] ‘tea’ or [xkafi} ‘coffee’ have voiceless aspirated stops in prevocalic position (also note that many loans will inflect with the English plural allomorph [-z], a voiced obstruent not in the inventory). In addition, there are several other native exceptions to the process. Rigsby (1986:134-138) gives the following list of palatalized coronal stops which are :37aspirated The claim here is that palatalization at the plosive release has aspiration superimposed on it. It could be argued that if /xi is normally phonetically realized as [ç], then a possible phonetic interpretation of the stop plus realease is [tc],which is underlying /tx]. While this is most likely the diachronic explanation for these forms, it is unlikely as a synchronic explanation, as /x/ surfaces prevocalically as the voiced [j]. Whether this U]should be considered a secondary articulation or a separate segment is unclear from the data.

110 (55) Exceptions to voicing a. [thiaks] ‘net float’ [thal’tJ ‘be good looking (of a house) (vi)’ [thiaj tXW1 ‘thunder’ [tl1uk] ‘dusk’ [sit”e:wa] ‘trade’ (vi) [sit”e:x”s] ‘change’ (vt)

As Rigsby points out, the velar fricative [xl is typically phonetically fronted, nearly to the point of being palatal. This segment does not occur in prevocalic environments, and so he posits underlying forms with the sequence /tx-/, along with a rule that converts the underlying fricative to the palatal glide [j] prevocalically. When the oral articulation for the fricative is lost, its secondary feature of palatalization remains; cf. discussion of the gliding rule in chapter 4. Additional forms, however, pose a different sort of problem. Forms with aspirated stops as in (56) are historically derived from [t]+[x’’] clusters, as are still found in Nisgha as [tJ+[g] clusters.

(56) Gitksan Nisgha [thun] [tgun] ‘this (one)’ [thust] [tgust] ‘that (one)’ [t”wa] [tg”a] ‘glass, quartz crystal’ [thwa:lixs_] ‘mix, stir’ (vt)

Parallel to (47) above, in these cases, when the velar fricative was lost historically, its secondary feature of labialization was retained. Since labialization is incompatible with a following [u] (Rigsby 1986), this feature is lost entirely in these contexts. Finally, there are forms which can be considered truly exceptional, in that there is no historical change that would have derived them, where the cognates in Nisgha indicate that the stops should be voiced. Such forms are found in (57):

ill (57) Gitksan Nisgha [t”o:jip} [do:?lipj ‘cook by steaming in earth-oven’ (vi) [tho:st] [do:st] ‘belongings’ [sitho:q] [sido:q] ‘coax, deceive, lead on’ (vi) [giphajkwhl [gibajk’] ‘fly (vi, sg)’

One possible analysis of the aspirated stops is to simply assume that they constitute a separate stop series in the inventory of the language, yet maintain a low functional load. This is essentially the approach taken by Sasama (1995) for Coast Tsimshian. One way of implementing this analysis is to assume that these exceptional aspirated stops are laryngeally specified as [sg], as in (48). This synchronic analysis may (55-56) or may not (57) have roots in diachrony.

(58) [-son]

Lar

[sg]

Next, the question of why voicing does not affect these stops must be considered. If the aspirated stops were to be voiced, the result would be a a consonant that is [voice, sg]; in other words, a breathy voiced consonant, or voiced aspirate, as is found in many languages of the world (Ladefoged & Maddieson 1994). However, this segment type is not found in Gitksan (and would, incidentally, be another instance of “doubling up” of laryngeal features).

4.3.2.3. Fricatives

Since aspirated stops were shown to be [spread glottis], we next turn to the behavior of fricatives. Fricatives do not undergo the prevocalic voicing assimilation that stops and affricates do:

112 stop uvular are (Ohala fricatives lenition voicing. characterized

[ican-rcanJ 38

and

It

4.3.2.4.

stops,

(60)

discussion,

The blocks

section

a

[+voice]/[+son].

(59) was The There derived

voicing,

structural

fricatives. representation voiced

&

a shown

of idea

voicing

In

So e:q lit’ saks

are

claims

Summary constraint is Riordan voiced ‘trees’ from

an a this here

a far, by

stops

see fashion normal

exception above phenomenon the uvular uvular

(Rigsby

Lar

[sg]

there

that

*lce:q

*1 3 it

description

While *zaks assimilation.

is [-son]

1979, Vaux because feature

that

can of

or

for

fricatives

[+cont] that stops.

stops

similar has Ohala allegro

to

Blocking

(1998),

Building the affricates,

1968:9).

be fricatives of

the

[sgJ, prevocalic been

to formulated will

the former Since 1983). voiced generalization

tempo

to difficulty in primarily

be

are

The

as

no

‘foam,

‘ball’ ‘to Presumably

while that

mentioned this

on

would

well a

This

are specified

variant simple

uvular

leave

idea the

process

plosive

above

rule that phonetically

in

assumed blossoms’ because

leaves

as that

maintaining discussion

fricatives.

(p1)’ then of

that

explanation Howe here,

this will the

is

that

for for

only

voicing voiceless

fricatives be variable, of voicing voiced

but

block

to

the analyzing

their as

the (2000)

characterized involves Rigsby

not be

voicing

found

laryngeal of uvular

uvular

fails pursued. voicing patterning

is specified and

do

for

fricatives blocking due

(1968:9, for not

ultimately to

during

in this fricatives

stop. the

to similar

voice

apply

(60)

[-cont], on

different

by feature fact,

with

exceptional underlyingly An stop

1986:154)

fricatives 39 .

laryngeal are a 8: in

not to as example fricated

stops

aside closure phenomena

Gitksan. [sg]

voiced, ejectives,

constraints [sg],

tied such with

explains

is from to in release

is

and

features

well-motivated; nature and the There /qanl

a for

respect that

posterior the

regular

aspirated that —

only in

that [cg], ones period, —>

is

of

statement Oowekyala.

to

the [-cont] this

a so these [Gan-Gan]

postlexical the

process aspirated that prevocalic consonant

ones

far, feature

are stops,

voiced latter

lenite 113

that this

not for

of

-) of two are assumed to be specified for [sg]. It is hypothesized that the failure of these segments to voice is due to a restriction on multiple laryngeal features being specified on a single [- son] root node.

4.4. Conclusion

The nature of the laryngeal features in Gitksan were discussed in this chapter. First, the dissimilatory and assimilatory effects that were found with the glottalized consonants in chapter 2 were discussed in more detail. Much like in chapter 3, these effects were modeled with weighted constraints. In addition, the actual agreement itself was modeled with the Agreement by Correspondence framework. The learning algorithm was run, and it was shown that acceptability scores for output forms correlated with their degree of attestedness. The phenomenon of plosive voicing was also discussed, as well as cases where certain plosive types, as well as fricatives failed to voice. This blocking phenomenon was attributed to these segment types being specified for other laryngeal features.

114 Chapter 5: Dorsal and Guttural Patterns

5.1.Introduction

This chapter explores the behavior of the velar, uvular, and laryngeal consonants in Gitksan. It is argued that the patterning of the uvular and laryngeal consonants in Gitksan defines a distinctive “guttural” class. This class typically includes uvulars, pharyngeals, and laryngeals (Hayward & Hayward 1989, McCarthy 1994), though in some languages laryngeals are placeless, and do not participate as gutturals (Rose 1996). It will be shown here that while Gitksan does not have pharyngeal consonants, it still maintains a class of gutturals including the laryngeals and uvulars, and that this class is phonologically distinct from, but overlaps the class of dorsals, which includes the velars and uvulars. The evidence for overlapping dorsal and guttural classes comes in many forms. Drawing on the results presented in chapter 2, it will be shown that there is an OCP-D0R effect active in the Gitksan lexicon. In order to account for this effect, the OCP-constraints employed in chapter 2 will be used. It is first necessary to establish this basic effect before the more complicated facts of gradient dorsal agreement are discussed. It will be shown that while dorsal-dorsal sequences are dispreferred, when two dorsals do happen to occur in sequence, they tend toward place agreement: velars tend to occur with velars, and uvulars tend to occur with uvulars. Parallel to the case of the gradient glottalization agreement discussed in chapter 4, the gradient dorsal agreement will be analyzed by means of the “agreement by correspondence” framework (Hansson 2001, Rose & Walker 2004), whereby correspondence is established between pairs of consonants in outputs, and these correspondent pairs are forced to agree with respect to some feature. The overall result of this gradient agreement is a tension between velar and uvular segments. In contrast, the cooccurrence restriction on the set {uvular, laryngeal} suggests an OCP effect, which is evidence for a guttural class. In addition to these gradient lexical effects, there also exist categorical effects which provide evidence for the guttural class. One such process is vowel lowering. Both laryngeals and uvulars lower vowels in certain contexts. In contrast, velars fail to lower

115 vowels. This evidence, along with the cooccurrence pattern, points to a representation for velars with [dorsal] as the place feature, for laryngeals as [pharyngeal], and for uvulars as simultaneously [dorsal] and [pharyngeal]. This feature representation derives both the “dorsal” type of classing responsible for the OCP effect, and also the “guttural” type of classing responsible for vowel lowering. This chapter can be summarized as follows: Section 2 discusses the patterns that dorsal consonants display over the Gitksan lexicon. These patterns include the general OCP effect briefly outlined in chapter 2, as well as a gradient agreement pattern. Section 3 presents evidence for a guttural class in the language. In order to define constraints on agreement, arguments are also made for the feature [pharyngeal]. These arguments are based primarily on vowel lowering, which is also outlined in this section. Section 4 presents an analysis of the lexical patterns. This includes an analysis for the OCP effects parallel to the analysis put forward in chapter 2, as well as an analysis for the agreement pattern. The agreement patterns are then analyzed by means of correspondence constraints. Section 5 concludes.

5.2. Lexical Patterns

In this section, the results from chapter 2 will be reviewed, and it will be shown that there exists an OCP-D0R effect active in the Gitksan lexicon. This effect was analyzed using ID- PLACE and the OCP constraints employed in chapter 2. It will be shown here that there is a gradient uvularity agreement effect in place, such that within the general context of place disagreement, when dorsals cooccur, they tend to agree in sub-dorsal place of articulation: velars tend to cooccur with velars, and uvulars tend to cooccur with uvulars. In order to capture these results, additional constraints on consonant pairs that share values for the feature [pharyngeal] will be introduced. The table in (1) provides an overview of the relevant patterns exhibited by dorsal consonant pairs in Gitksan. While velar-uvular pairs are significantly underrepresented in transvocalic (CXVCY)contexts, velar-velar and uvular-uvular pairs are both significantly overrepresented. This pattern is reproduced in long-distance contexts (Cx...C..

116 (1) Dorsal agreement: long and short distance

cxvcy cx.. .c.. Velar-Uvular 0.57* 0.66* (10/17.5) (15/23) Velar-Velar 1.40 * 1.49* (13/9.3) (18/12) Uvular-Uvular 1.46 * 1.24* (12/8.2) (20/16.1)

The question can now be raised as to how this gradient pattern relates to more productive processes. In other words, the question is whether we find any instances of dorsal harmony crosslinguistically, and if so, does it operate in a way that would be consistent with the patterns found in Gitksan? The answer is in the positive. Hansson (2001) shows that there are indeed languages that exhibit consonant harmonies based on dorsal features. One such categorical process is found in Misantla Totonac (MacKay 1994, Hansson 2001). In Totonac, there exists a type of uvular assimilation which operates under adjacency, whereby any /k/ which precedes /q/ also becomes /q/ (and uvulars are further subjected to simplification in some instances, though this varies between Misantla Totonac and Tlachichilco Tepehua). The language also exhibits a long distance uvular harmony, where any velar stop /k/ that precedes a uvular stop /q/ will become uvular. This process is limited to derivational prefixes.

(2) Misantla Totonac uvular harmony /ut maka-$qat/ [?t maqá5qt] ‘s/he scratches X (with hand)’ /ut maka-paS/ [?t makapa$] ‘s/he bathes her/his hand’ /min-kk-paq?/ [mINqqpa?] ‘your shoulder’ /min-kk-t5-nj/ [mIkkt$n] ‘your shoulder’ /maka-1uqwan-la(i)/ [maqa+6qwa+J ‘s/he tired X’

A very similar pattern is found in Tlachichilco Tepehua (Watters 1988, Hansson 2001), which is illustrated in (3).

117 (3) Tiachichilco Tepehua uvular harmony /?uks-laqts’inl [?oqslaqts’in] ‘look at Y across the surface’ /mak-t$aq’a:-j/ [maqt$a?a:j] ‘X washes hands (impf.)’ /lak-tiq’i-+/ [laqtSe?e+] ‘X broke them (perf.)’

In these cases, there is a categorical process, attested to by alternations, that militates against mixed (where “mixed” = disagreeing in uvularity) dorsal pairs. While the uvular harmony of Gitksan is only a gradient condition holding over roots, we can see a clear relation to the categorical, word-based processes like those found in Totonac and Tepehua. There are other cases of uvularity agreement that do not result in productive alternations; i.e., are limited to roots as a morpheme structure constraint. MacEachern (1999) and Hansson (2001), citing data from De Lucca (1987), also describe a similar state of affairs in Bolivian Aymara. In this dialect of Aymara, there is a prohibition on cooccurring velars and uvulars within a root:

(4) Bolivian Aymara (Hansson 2001:95) qeiqa ‘document’ qhatSqha ‘rough to the touch’ q’enq’o ‘rough (ground)’ crooked’ qhapaqa ‘wealthy, rich person’ kiki ‘similar, identical’ k’usk’a ‘common’ k’ask’a ‘acid to the taste’ k’iku ‘wise’ (obsolete)

Bolivian Aymara exhibits a categorical restriction as a morpheme structure constraint. A similar case is found in Ineseño Chumash (Applegate 1972, Hansson 2001:97), which exhibits a similar pattern. In this particular language, however, there are frequent cases of uvular-velar pairs, though as Applegate points out, many of the morphemes in which these occurrences are found could have possibly been morphologically complex at an earlier diachronic stage. Thus, it is likely that the gradient dorsal agreement found in Gitksan is similar to Ineseño Chumash, or something of a gradient version of a morpheme structure constraint like that in (4).

118 In order to formalize the gradient uvularity agreement in Gitksan, it is necessary to first determine what this agreement is representationally based on. In other words, what is the feature that uvular-uvular or velar-velar pairs agree with respect to? It will be argued in the next section that the relevant feature distinguishing uvulars from velars is the feature [pharyngeal].

5.3. Arguments for Dorsal vs. Pharyngeal Place

The following illustrate (near) minimal pairs for uvular and velar stops (including the voiced vs. ejective series):

(5) Velar/uvular (near) minimal pairs gat ‘man’ Oats ‘pour’ k’ap ‘really, certainly (verb proclitic)’ q’ap ‘piece, part, relative’

Since the uvulars and velars are contrastive in the language, there obviously must be some way for the grammar to distinguish these consonants. The ultimate claim here will be that velars and uvulars are both specified as [dorsal], but that uvulars have an additional specification for [pharyngeal] place (Cole 1987, McCarthy 1989, 1994). One way of distinguishing velars from 4uvulars is to adopt either Halle’s (1989, 1992, including Halle et al. McCarthy’s° 2000), or (1989, 1994) approach to guttural representations. For Halle, the guttural node dominates both the laryngeal node and the tongue root node, as represented in (6). The Place node dominates features for oral articulators such as the lips [LIPS], tongue blade [TBL], and tongue body [TBD]. Dorsals would be specified with the typical Place feature [DORSAL] (dominated by [TBD]), and uvulars would also be specified as [DORSAL], but with the added Tongue Root feature [RTR] (or [RADICAL], under the Halle et al. 2000 approach).

40 There have been many approaches to distinguishing velars from uvulars. One is to view the velars as [+highj and the uvulars as [+backj, with the laryngeals and pharyngeals as [+IowJ (Chomsky & Halle 1968:307). This type suffers from inconsistencies in feature defmitions/uses and phonetic implementation. This is discussed further in section 3.1, and specific arguments can be found in Kenstowicz & Kisseberth (1979), Keating (1988), and McCarthy (1989, 1994).

119 (6) Halle (1989, 1992), Halle et al. (2000) Root

Guttural Place

Laryngeal Tongue Root LiPs TBL TBD A [sgj [cgj, etc. [ATRI [RTR]

Under this view, the guttural consonants (uvulars, pharyngeals, laryngeals) are predicted to pattern with other laryngeal features, such as [spread glottis] or [constricted glottis]. This is not the case, though, in Gitksan. For instance, while uvulars and laryngeals lower vowels (section 5.3.1. below), ejectives and glottalized sonorants fail to do the same. Cole’s (1987) and McCarthy’s (1989, 1994) model makes predictions that better fit the Gitksan data in that gutturals are not predicted to interact with tongue ’4root and laryngeal features. These models places the feature [pharyngeal] as a daughter of the Place node, but separate from the Oral place node:

(7) Root Laryngeal- Place [voice] [cgj [sg] Oral

[lab] [cor] [dor] [phar]

Under this view, dorsal consonants, including velars and uvulars, are specified with the place feature [dorsal], and uvulars are specified as both [dorsal] and [pharyngeal]:

(8) Place

Oral [phar]

[don

41 This configuration forces tongue root features such as [RTR] to be located elsewhere in the geometry, such as on a separate Vowel Place tier (Odden 1991).

120 This is relatively uncontroversial, and assumes that velars in inventories with gutturals behave exactly like velars in inventories that lack gutturals.

(9) Velars (k, g, x, k”, X”) Place

Oral

[don

While the representation for dorsals and uvulars was for the most part straightforward, the representation of laryngeal consonants, on the other hand, is a bit more controversial. Under McCarthy’s model, laryngeals are specified as being [pharyngeal], as are all other members of the guttural class. This feature is claimed to be an oro-sensory feature (McCarthy 1989, 1994)which is defined by the region in which uvulars, pharyngeals, and laryngeals are produced (as opposed to an active articulator).

(10) Low gutturals (? h) Place

[phar]

The difference between this model and the one in (7) is that while gutturals bear the feature [pharyngeal], this feature is dominated by the place node, and does not dominate other tongue root features, nor crucially, laryngeal features. Under this view, the dorsal consonants, including velars and uvulars, are predicted to pattern together since they share a specification for [dorsal]. This evidenced by some lention facts of the language; for instance, all of the dorsal fricatives ([x x” xl)undergo the ‘gliding’ rule (Rigsby 1986). In post-tonic, intervocalic position, the dorsal fricatives become the glides [j, w, h]. This is illustrated in (11), where the singular nominals are contrasted with their possessed counterparts:

121 (11) Dorsal gliding wa:x ‘paddle’ wa:jin ‘your (sg.) paddle’ muxw ‘ear’ muwin ‘your (sg.) ear’ nox ‘mother’ nohon ‘your (sg.) mother’

Further evidence comes from reduplication. Rigsby (1986) notes that dorsal stops spirantize in the coda of CVC- reduplicants. Dorsal [k k’’ q] (or their voiced counterparts) become [x

X” ]‘ respectively. Ejectives (as well as glottalized sonorants) deglottalize in reduplicants. This is illustrated in (12).

(12) Reduplicative lenition ?akst ?ax-?akst ‘be wet (intrans)’ ts’ak’ cix-ts’ak’ ‘dish’ ts’i:k’ cix-ts’i:k” ‘to leak’ doq’ da-doq’ ‘to be deaf naq’ na-naq’ ‘skirt’ coq cka-joq ‘to camp’

Furthermore, labial and coronal stops do not undergo this lenition in reduplicants, as (13) shows that the stops surface faithfully in the same position.

(13) Non-lenition of labial, coronal stops chap cip-cap ‘to make, do’ +ap +ip-+ap ‘be deep’ mitxv mit-mitx’’ ‘to be full’ hadixs hat-hadixs ‘to swim’

Other dorsals such as uvulars, which cross-linguistically tend to phonologically pattern with pharyngeals and laryngeals, are specified as [dorsal] and as [pharyngeal] place. This is true in Gitksan, as evidence presented below, including cooccurrence restrictions and

122 vowel lowering attests. This natural classing implies that the representation of uvulars is as in (14), with specifications for both the [dorsal] and [pharyngeal]features.

(14) Uvulars (q, G, x) Place

[don [phar]

This type of configuration captures the overlapping dorsal and pharyngeal classes quite well: velars and laryngeals typically have little in common, and thus are represented with different place features ([dorsal] vs. [pharyngeal]). Uvulars, on the other hand, share some characteristics with velars, and other characteristics with laryngeals (such as vowel lowering, discussed in section 5.3.1). Thus, uvulars are specified as being simultaneously both dorsal and pharyngeal (with no distinction between primary and secondary articulation being drawn for Gitksan). Evidence for the guttural class in Gitksan comes from cooccurrence restrictions themselves. When only uvular and laryngeal consonants are isolated, it can be observed that there is a gradient (i.e. non-absolute) cooccurrence restriction against these sounds aggregated as a class. For instance, when testing for guttural {uvular, laryngeal} vs. dorsal

{velar, uvular} restrictions, the guttural-guttural pairs in transvocalic contexts have an OlE value of 0.67 (16/23.8, p <0.05), which is moderately underrepresented. Since it has been shown that the uvular-uvular pairs are themselves overrepresented (due to the gradient agreement phenomenon) within the larger dorsal macro-class that overlaps with the guttural class, the total identity pairs need to be excluded (rather than included as when the uvular uvular and velar-velar pairs were observed). These identical pairs must be removed, otherwise the results would be obscured by the total identity exemption, which sets the assimilatory effect apart from the dissimilatory one. The 0.67 OlE value was determined in this fashion (with identicals removed), which is parallel to how values for the dorsal-dorsal pairs were determined in Chapter 2. This measure includes any combination of uvulars or laryngeals in either 1C or 2C position. In other words, this includes uvular-uvulan,laryngeal laryngeal, and uvular-lanyngeal pairs. If the class of gutturals is analyzed in terms of its component parts, i.e. broken down into uvulars and laryngeals, the significance of any under-

123 or overrepresentative values for OlE disappears, aside from the laryngeal-uvular pairs (perhaps due in part to the small number of examples available):

(15) Guttural sequences Uvular Uvular 0.82 (n.s.) (6/7.3) Laryngeal 0.50* 1.22 (n.s.) (7/14) (3/2.5)

Despite this, the direction of the numbers here could be taken as suggestive, as uvular-uvular pairs and laryngeal-laryngeal pairs have values greater than one, and the mixed laryngeal uvular pairs have a value significantly lower than one. This is what would be expected if the uvulars and laryngeals formed a natural 42class If uvulars and laryngeals did not form a natural class, we would expect O/E values .for the mixed pairs to be near 1. While the uvular uvular and laryngeal-laryngeal cells are the ones with the most similar sequences, it has already been demonstrated that there is an assimilatory pressure affecting uvulars, and the numbers for laryngeal pairs are extremely low. Thus, the laryngeal-uvular underrepresentation in (15) appears to be a genuine OCP effect, and the 0.67 OlE cited in the paragraph above attests to the overall dissimilatory effect for guttural-guttural pairs. Such an effect is categorically found in Arabic, where root-internally, tier-adjacent gutturals are banned (McCarthy 1994). This suggests that in addition to the OCP[place] constraints that have already been introduced, there is an additional constraint prohibiting guttural-guttural pairs: OCP-PHAR.

5.3.1. Guttural Lowering

In addition to the evidence from gradient lexical effects, there is also strong categorical evidence for the laryngeals and uvulars behaving as a class of gutturals. In Gitksan, as in Nisgha and Coast Tsimshian, uvulars and laryngeals trigger the lowering of vowels that are immediately adjacent (see also Rigsby 1967, 1986 for the generalizations in Gitskan, Tarpent

42 Actually, the diagonal cells, or at least the uvular-uvular cell would be expected to have low OlE values due to the combined OCP-D0R and OCP-PHAReffects. However, this particular cell is inflated due to the dorsal agreement discussed in section 2.

124 1983, 1987, Thompson 1983, Shaw 1987 for vowel lowering in Nisgha, and Dunn 1979, 1995 for similar facts in Coast Tsimshian). For reference, the vowel inventory of Gitksan is given in (16):

(16) Gitksan vowel inventory ii: uu: e: o: a a:

Vowel lowering can be illustrated most clearly in cases of reduplication. In Gitksan, reduplication typically involves a CV(C)- or CVx- copy of the base, but where the vowel is of predictable quality. In most environments, the vowel that surfaces is [i], as illustrated in (17).

(17) Fixed vowel reduplication saksx”’ six-saksx” ‘to be clean’ ckam ckim-&am ‘to cook, boil (trans.)’ lak li-lak ‘to be crooked’ gwalkw g”il-g”alk” ‘be dry’ lap lip-lap ‘be deep’ gup gip-gup ‘to eat (trans.) m’asx” mis-m’asx’’ ‘to sting (trans.)’ daw di-daw ‘ice, to freeze (intrans)’

However, when preceding or following a uvular, the vowel surfaces as [a]. Since there is no input vowel to be faithful to, this vowel is determined entirely by markedness constraints combined with Base-Rediplicant correspondence constraints.

(18) Vowel lowering in reduplicant dzoq dza-dzoq ‘to camp’ Getx” Ga-GetX” ‘to be difficult, be expensive’ doq’ da-doq’ ‘to be deaf’ lo:q la-lo:q ‘to be early’ t’oq da-t’oq ‘to grab’

125 Go:t Ga-Go:t ‘heart’ GOS GaS-GOS ‘to jump’

So far, it has been assumed that laryngeals are specified with a place feature, and that the feature is [pharyngeal]. This is also not a trivial assumption. There have been claims in the literature that laryngeal consonants are placeless in some languages (Clements 1985, Sagey 1986, Steriade 1987, Bessell & Czaykowska-Higgins 1992, Stemberger 1993, etc.), which would imply that they do not pattern with uvulars and pharyngeals, which are [Pharyngeal]. There have also been claims to inventory structure being a determinant of laryngeal specification. Rose (1996) has suggested that when a pharyngeal segment is present in an inventory, laryngeals will be [pharyngeal]. However, when no pharyngeal is present, laryngeal consonants will be placeless. It is important to test these hypotheses with regard to the features for laryngeals in Gitksan, as the constraints on gutturals will necessarily need to reference these segment types. What pertains in Gitksan is this: the language does not have any pharyngeals in its segment inventory, which would imply (according to Rose 1996) that the Gitksan laryngeals are placeless. However, this cannot be the case, as laryngeals participate, along with the uvulars, in the lowering of vowels:

(19) hets has-hets ‘to send’ ?os ?as-?os ‘dog(s)’

This indicates that the laryngeal consonants in the language have a place specification, and that this place is one that is shared by the uvulars. Thus, the laryngeals are [pharyngeal], contra Rose (1996). In contrast to the behavior of the laryngeals and uvulars, the velars do not lower vowels (velar fricatives only occur in post-vocalic position):

(20) No lowering by velars gat gi-gat ‘to be born, to hatch (p1)’ gin gi-gin ‘to give (p1)’ gida gi-gida ‘to ask (p1)’

126 gwalgwa ‘to be thirsty (p1)’ gi:s gis-gi:s ‘to be wrong, miss (p1)’ gup gip-gup ‘to eat (trans) (p1)’ ga?a gix-ga’?a ‘to see, look at (p1)’ muk’’ mixmukw ‘purple (p1)’ bil’ust bix-bil’ust ‘star (p1)’

This suggests that the velars are not part of the guttural class, and hence are not specified as

[PHARYNGEAL]. Returning to the point made about lowering in reduplicative contexts, it is important to note that these contexts can derive vowels that either precede or follow gutturals, and that in either case the vowel surfaces as a low vowel. This suggests that the direction of the process is not limited to one side. This contrast between the fixed high vowel and the appearance of the mid and low vowels suggests that outer domains of morphology, such as reduplicants, are locations whereby the least marked configuration is allowed to emerge (McCarthy & Prince 1994). Thus, the low vowel emerges when adjacent to gutturals because this is the unmarked consonant-vowel configuration; elsewhere, the high front vowel emerges because it is the unmarked vowel in the language. This phenomenon is not limited to reduplicants. Vowel lowering in affixes is nearly identical to lowering in reduplicants. The final vowels of prefixes will lower if adjacent to a uvular or laryngeal in the stem. (a) examples show the unmarked (or rounded) cases; (b) examples illustrate instances of lowering.

(21) ii- ‘DEF’ a. li-ts’aqt ‘the tip of it’ ii-ts’ew’+ ts’imuwij’ ‘the interior of my ear’ b. ia-Gasipdi:t ‘their bones’ +a-habi+‘Ianckam ‘the lid of the kettle’

(22) Si- ‘INTRANS’ a. si-ma:j’ ‘pick berries’ si-ts’aq’ ‘dig, gather clams’ b. sa-?is ‘pick soapberries’

127 Sa-GasX ‘dig wild rice’

Thus, like reduplicants, affixes display vowel lowering properties. Finally, vowel lowering takes place in compounds. While the process of compounding is less productive than the other areas of Gitksan morphology, compounds do provide another context in which morphological lowering takes place. Compound heads ending with a vowel such as Isp-/ ‘liar, den’, /xso-/ ‘water, river, stream, fluid, juice’ (Rigsby 1986:67) display this property:

(23) anda-, andi-, 43andu- ‘container for something’ a. ?andi-+gu:+x9 ts’u:ts’ ‘bird’s nest’ ?andu-wo:t’ ‘pocket’ b. ?anda-hawil ‘quiver’ ?anda-?is ‘bladder’

(24) Compounds with [i] a. xsi-m’o:t’ixs ‘milk’ (one consultant: milky) b. xsi-j’ans ‘dew’ c. xsi-gunja?a ‘Salmon River’ d. xsi-txemsim ‘Nass River’

(25) Compounds with [a] a. xsa-?ando:?o ‘Bulkley River’ b. sba-?ax”t ‘porcupine’s den’

Compounds, while arguably losing morphological productivity in the language, still provide alternations of low and high vowels in the context of gutturals vs. elsewhere, exactly like the other types of morphological lowering mentioned above. Furthermore, many instances of compound formation in Gitksan, such as those above, rely on the affixation of a

‘ It is interesting to note that some speakers do not display variations in vowel quality in this prefix; rather, the final vowel of the prefix always surfaces as [i]. Interestingly, there are also words that have optional pronunciations with lowering: huw’andi: - haw’andi: ‘not yet’.

128 lexical bound root. This root behaves phonologically like an affix, and will be treated alongside the other affixes in the analysis of vowel lowering. A claim in the literature is that there is a morpheme structure constraint operative in Gitksan roots whereby gutturals lower vowels one height (Rigsby 1986). In this way, high vowels would be lowered to mid, and mid vowels would be lowered to low vowels.

(26) Root internal patterns of uvulars and midllow vowels baX ‘to run (sg.)’ Ga:p ‘to scratch (trans., sg.)’ Galan ‘after’ Golix ‘scalp’ Go:t ‘to be empty’ GOS ‘to jump’ Gen’ ‘to chew’ oenx ‘to fall (of tree) (intrans)’ Ges ‘hair’ ?e:q ‘coho salmon’ le:q ‘to be worn’ +o:q ‘to be early ‘squawfish’ me:x ‘to be sour (of milk, etc.)’

(27) Root internal patterns of laryngeals and mid/low vowels ?a:da ‘to be proud’ ?aq ‘mouth (outer opening), lips’ ?a:t ‘ashes’ ha ‘air’ ha?aks ‘bucket’ ha:t ‘intestines, guts’ ?e:q ‘coho salmon’ ?os ‘dog’ helt ‘to be much, many’ hon ‘anadromous fish, salmon (but also includes steelhead trout)’ ho:ni ‘ankle’

129 There is definitely a trend toward this state of affairs in the Gitksan lexicon. However, it is not exceptionless. In fact, the number of exceptions, especially with respect to the laryngeals, warrants an investigation into the cooccurrence patterns of gutturals and vowels. For instance, there are many cases of gutturals adjacent to high vowels that can be derived by sound changes, or through reduplicative processes that derive gutturals from other underlying consonants. Rigsby (1986:186) notes that “some initial /hls (those that appear before [i] and [u]) probably descend from initial /x/ and /x’”/, respectively”, though no evidence is presented to support this. There is, however, other evidence to suggest that another diachronic pathway is responsible for some [h]s. There are several exceptional forms that have cognates in Coast Tsimshian (forms are from Dunn 1995, where the page number and dictionary entry number are indicated).

(28) Gitksan Coast Tsimshian winGis ‘brain’ [winGaws] pg. 111, 2121 ?i:s ‘necklace’ [juits”k9 pg. 118, 2241 ?i:w’xt ‘man (p1)’ [?ju:thaj pg. 118, 2250 ?ubin ‘to be pregnant’ [wa:jbi] pg. 106, 2042

It is possible that the glides in the Coast Tsimshian cognates vocalized in Gitksan, though it is predicted that the labiovelar glide in Coast Tsimshian [winGaws] would correspond with an [u] instead of an [iiin Gitksan. In addition to possible diachronic changes, there is also a synchronic process in the language that derives surface [hI from an underlying labial sonorant. The default vowel found in reduplicants is the high front vowel [i]; however, the predictable or “fixed” vowel will show up as rounded if followed by a labial consonant. Further, there is a cooccurrence restriction in reduplicants that prohibits a sequence of a labial consonant followed by a rounded vowel (where the round vowel is triggered in the reduplicant by a following labial consonant). The result is a replacement of the labial consonant by [h], or more accurately, the process can be viewed as debuccalization (i.e. the loss of place features):

130 (29) Labial - [h] reduplication wilp hu-wilp ‘house’ wa hu-wa ‘name’ max” hu-max”' ‘burst out laughing’ m’al hu-m’al ‘canoe’

It is tempting to claim that this is not total debuccalization, as place features could be retained in a very limited sense: the place feature [labial] could be copied and preserved in the reduplicant, but realized on the following vowel instead of the consonant. In other words, it could be the case that [labial] is preserved by base-reduplicant correspondence (between i of the base and the reduplicant vowel). This, however, is not correct, as CVC- reduplicants illustrate that the assimilation of the vowel must be to a following labial consonant:

(30) Rounding assimilation blocked mitxv mit-mitx”' ‘be full’ ma:xwsx mis-ma:x”’sx” ‘be white’ m’ats mis-m’ats ‘to hit, strike’ w’in win-w’in ‘tooth’

In every case of a labial consonant followed by round vowel in reduplicants, what surfaces is a derived [h] followed by a high round vowel. There is another reduplicative context where [hJ surfaces preceding a high vowel; this time, as a correspondent of U]in the base. This is due to a cooccurrence restriction on the palatal glide followed by the high front vowel (that is, a ban against 44[ji]). Again, this is likely a case of debuccalization.

(31)j - [h] reduplication a. lu: jaltx”' lu: hil-jaltx”' ‘return, come back (intrans)’ b. jo?oxs hi-jo?oxs ‘to wash face, hands, etc.’ c. jal hilijal ‘to lie, tell a lie’

Although this generalization is not entirely productive; there exist reduplicants that faithfully copy the base palatal glide, such as jim/jim-jim ‘smell (trans)’ and jats/jis-jats ‘to hit’.

131 This is immediately transparent in (3la-b), though sound changes have made (3ic) less identifiable as a reduplicant. In both cases of reduplication, it can be seen that [h] preceding a high vowel is derived from an underlying (non-guttural) sonorant. One & Bricker (2000) present a similar case in Yucatec Maya, where there exist a class of [his that behave as gutturals (i.e. specified as [pharyngeal]), and a class of historically derived laryngeals which behave as if they are placeless, and not as gutturals. Orie & Bricker claim that the two types of [hi in Yucatec have representational differences, and that these differences boil down to placeless vs. pharyngeal laryngeals. Following One & Bricker (2000), the approach here assumes that these (historically) derived laryngeals are placeless, and not specified as [pharyngeal]. Rather, they are only specified for laryngeal features (in this case [sgj). The representation for this placeless IhI is shown in (32a), which

can be contrasted with the representation for a guttural 1W,shown in (32b):

(32) a. Placeless 1W b. Guttural Ihl RT RT

Laryngeal Place

[sgj [pharyngeal]

It was claimed in chapter 4 that fricatives are specified as [sg]; the same is true of 1W. Hence, (32b) subsumes (32a) in that in addition to a Place node, there is also a laryngeal node specified for [sg]. Since debuccalization is responsible for the non-lowering behavior of these derived [his, then the argument boils down to this: the feature [pharyngeal] can be inherited through Base-Reduplicant correspondence (as in ex. 19), but it will never be inserted simply to satisfy a putative markedness constraint that demands that all laryngeals be specified as [pharyngeal]. This is evidenced by the cases of reduplicant [hj in (29) and (31), which are not in BR correspondence with a specification of [pharyngeal] in the base segment. Despite these synchronic and diachronic explanations for exceptional sequences of guttural plus high vowel, there are also a substantial number of exceptional cases that have no obvious explanation. It is assumed here that these cases also consist of the non pharyngeal [h]s, which actually falls out from standard assumptions about markedness and

132 the emergence of the unmarked (McCarthy & Prince 1994). It was claimed above that there are two representational types of [hi in reduplicants. Since the general architecture of OT predicts that reduplicants (which are not protected by Input-Output faithfulness constraints) can never exhibit more diversity in surface structures than non-reduplicant material (which is protected by Input-Output faithfulness constraints), then it should follow from this that laryngeals not specified for [Pharyngeal] should be possible in underlying representations. This also follows from the principle of Richness of the Base (Prince & Smolensky 1993 [2004]). Such cases are listed below in (33).

(33) Non-lowering [his: hi:+ux’” ‘morning’ hil’a ‘close, nearby’ hit’ ‘scar, heal’

Since this appears to be a static, root-level pattern, this potential morpheme structure constraint (and its exceptional nature) will not be discussed in the context of lowering in derived environments. The gutturals of Gitksan so far have run parallel to the patterning of gutturals in languages like Arabic. However, the uvular stops don’t pattern exactly as they do in Arabic. In Arabic, uvular stops do not pattern with the gutturals in lowering vowels, which forces McCarthy to provide an alternative representation for these segments. Unlike Arabic, the uvular segments in Gitksan pattern uniformly (i.e. uvular stops pattern with uvular fricatives as gutturals), as can be observed in the data presented above. Thus, there is no need to make any adjustments to their representation. Further, as will be shown below, the specific constraints employed in analyzing vowel lowering are markedness constraints that prohibit adjacent sequences of gutturals and vowels of varying height. Since these constraints are defined over adjacent sequences, there is no longer a need to represent guttural lowering as the result of autosegmental spreading (as assumed in McCarthy 1994 for Semitic).

133 5.3.1. Constraints on Guttural Lowering

This section will explore what the nature of the constraints on vowel lowering must be. Since the relationship between a guttural segment and an assimilated vowel is one of adjacency, I assume syntagmatic constraints that govern the weilformedness of adjacent segments. It is also clear from the morphological instances of lowering that both the laryngeals and the uvulars pattern together. Thus, the feature used to class the guttural consonants, [pharyngeal] (McCarthy 1994), will be adopted to account for this behavior (see also Prunet 1994).

(34) *PHARThGEAL/VHEIGHT: Segments that are adjacent and are specified for [PHARYNGEAL] and vowel height features incur a violation

Adopting this type of constraint is an alternative to autosegmental spreading. In his analysis of vowel lowering by gutturals in Semitic languages, McCarthy (1994) adopts a representation where the [pharyngeal] feature spreads from consonant to vowel.

(35) Guttural lowering (following McCarthy 1994) C y

[phar]

Rather than the autosegmental spreading of the [pharyngeal] feature, this analysis adopted here (based on the constraint in 34) instead views vowel lowering as the result of a constraint on sequential feature 45cooccurrence Obviously such a constraint is useless, since it will rule out not only sequences that are not lowered, but. any sequence of a [pharyngeal] consonant and a vowel. Thus, the constraint must be broken down into subconstraints defined in terms of specific vowel heights. In order to rule out certain vowel heights but allow others, the constraint must be further articulated. For the moment, it will be assumed that the constraints derived from (34) are in a stringency relationship (de Lacy 2002, 2004).

‘ While sequential feature prohibition such as in (34) could in principle result in autosegmental spreading, the actual feature itself that would have to be assumed to spread is in this particular case undesirable as both a consonantal and vocalic feature (see further discussion in the following paragraphs).

134 (36) *PHAPJMID>> *PHAPJLOW

This scale is a shorthand for the following, more detailed scale:

(37) * PHARI[+high,-low] >> * PHAR![-high, -low]>> * PHARI[-high,+10w]

Since the lowest member of this hierarchy is not essential (i.e. there are no penalties associated with sequences of gutturals and low vowels), this constraint can be removed (cf. de Lacy 2002, 2004, Gouskova 2003). This further simplifies the system, as the remaining two constraints can be reformulated with only a single vowel feature:

(38) * PHARJ[+high]>> * PHAR![-low]

This approach to consonant-vowel assimilation clearly rejects feature theories that treat [low] as equivalent to [pharyngealj (McCarthy 1994, Prunet 1994), and it also rejects the idea that pharyngeal consonants are specified as [+low] (Chomsky & Halle 1968). While both of these approaches would allow the assimilatory interaction between guttural consonants and low vowels to be viewed as either feature spreading or feature agreement (though not necessarily through correspondence), the constraints above do not. This approach avoids associating guttural consonants with the feature [+low] (as McCarthy 1994 points out, uvulars require a high tongue body for their articulation). It also makes clear the need for an interaction between guttural consonants and mid vowels, as is found in roots (see 26-27). If these static patterns are to be analyzed as the gradient glottalization agreement phenomenon was in Chapter 4, then there must be some features available for the grammar to distinguish guttural+mid and guttural+low sequences. [PHARYNGEAL] as a vowel feature is inadequate in this respect, as it would only characterize low (and not mid) vowels. Rather, the approach endorsed here places the “logic” behind guttural lowering (and potentially all C-V interactions) in the concrete phonetic domain rather than in terms of representational similarity or the spreading of features.

135 The type of scale in (36) is phonetically grounded, in the sense that there is an acoustic and articulatory relationship between the articulation of a guttural consonant and the height of a neighboring vowel. McCarthy (1994) has discussed the various phonetic properties of guttural consonants, and has noted that they are acoustically characterized by a high Fl locus. This is supported by numerous acoustic studies, especially works such as Alwan (1986), and Zawaydeh (2004). McCarthy states that “Fl is at the theoretical maximum in the case of laryngeals, close to the maximum for the pharyngeals, and higher than any orally articulated consonants in the case of uvulars” (pg. 196). Given a high Fl as one of the defining characteristics of gutturals, we can see how a transition to or from a vowel will also have an effect on F1. Since F1 is correlated with vowel height, then those vowels with the highest Fl (the low vowels) will form a smoother transition from or into a guttural than those vowels with a lower Fl (mid vowels), or the lowest Fl (high vowels). Thus, markedness disfavors vowels with lower Fl to be adjacent to gutturals, or to put it in terms of the formalism in (36), the configuration of [pharyngeal] consonants and high vowels is least preferred, the configuration of [pharyngeal] consonants and mid vowels is next least preferred, etc. Interleaving a faithfulness constraint like IDENT(VHEIGHT) within the constraints in (38) will then produce a set of possible inventories of guttural consonant/vowel configurations, each of which is outlined below. Importantly, interleaving IDENT(VHEIGHT) with the *PHARThGEAL/VHEIGHT constraints does not derive the vowel inventory of a language, simply because of the nature of *PHARThGEAL/VHEIGHT, which only derives statements about the height of vowels in the context of [pharyngeal] segments. Instead, this configuration derives the vowel inventory in a particular set of environments (particularly, adjacent to a guttural). At first glance, the morphological pattern of vowel lowering looks like a simple case of Emergence of the Unmarked (McCarthy & Prince 1994). For instance, while we are assuming that lowering of this type does not take place within roots (as there are plenty of cases of gutturals adjacent to mid vowels, for instance), we will see now how the dynamic pattern emerges. In order to derive this difference in morphological domain, there must be a faithfulness constraint (IDENT(VHEIGHT)) that is separated into two constraints: one for the root domain (Root-Faith), and one for the domain of outer morphology (Affix-Faith),

136 including functional affixes and lexical bound roots. McCarthy & Prince (1995) have claimed that Root-Faith>> Affix-Faith universally, and this is true for the data at hand. In order to block any kind of lowering root-internally, IDENT(VHEIGHT)-ROOTmust outweigh the entire set of markedness constraints. However, in order to derive the pattern of gutturals lowering vowels (with the only result being a low vowel), then IDENT(VHEIGHT)-AFFIXmust be weighted below the markedness constraints.

(39) Emergence of low vowels in derived environments

- 2 2 1 1 /+i-qa-[sipj-tiit/ IDENT(VH)-ROOT *PHAR!HIGH *PHAPJMID IDENT(VH)- H ‘their bones’ AFFIX a. +iGasipdiit -1 -2 b. +eGasipdiit -1 -1 -2 C. faGaSipdiit -l -l

(40) Emergence of low vowels in reduplicated environments 2 2 1 1 H /RED+Gos/ ‘jump’ IDENT(VH)-ROOT *PHAPJHIGH *PHARJMID IDENT(VH)-AFFIX a. Gas-Gas -l -1 b. GOS-GOS -2 -1 -3 c. GiS-Gos -1 -1 d.Gas-Gos -1 -1 -2

In cases of reduplication like that in (40), IDENT(VH)-ROOT prevents any alteration of the vowel height of the root (and by extension, prevents any kind of back-copying effects). The effect of this constraint being highly weighted is that vocalic root material will remain unchanged. This constraint rules out candidate (39a). Candidates (38b) and (38c) are ruled out by their violations of *PHAPJ/HIGHand *PHApJMID This leaves candidate (38d), with only a single violation of the lowly weighted *PHARIMIDand IDENT(VH)-AFFIX. Thus, only the markedness constraints will be relevant for the surface quality of this vowel. Since these constraints are arranged on the scale *PHAPJHIGH>>*pHApIMID,high vowels and mid vowels will be ruled out when adjacent to a guttural. This leaves candidate (39d) as the winner, as it has a low vowel in the reduplicant. This is an example of the Emergence of the Unmarked (McCarthy & Prince 1994).

137 This interleaving of markedness constraints and faithfulness constraints captures the emergence of guttural + low vowel combinations in morphological contexts. It should be noted here that while this exposition has focused on the difference between IDENT(VH)-ROOT and IDENT(VH)-AFFIX, both of these constraints need not be adopted to capture the facts. In order to capture the asymmetry between vowels in roots and vowels in derived contexts, all that is necessary is the special constraint IDENT(VH)-ROOT outranking the general constraint IDENT(VH),such that these constraints are in a stringency relationship (cf. Archangeli & Pulleyblank 2002 on Kinande, as well as de Lacy 2002, 2004). It is also worth pointing out that there is a fundamental difference in the treatment of C-V assimilation, which is handled with the set of markedness constraints above, and the consonantal agreement (which can be considered a type of assimilation), which is analyzed using the agreement by correspondence paradigm. It might at first be tempting to treat all cases of assimilation in the same fashion, but there is an immediate problem with applying the agreement by correspondence approach to guttural lowering: it opens up the range of correspondence much further than would be desired. If C-V are assumed to be in correspondence, which is what would be necessary for “pharyngeal” or “low” agreement, then this would imply that all consonants would be in agreement (i.e. not just obstruents, or plosives, etc.). That is, if C-V are correspondents, then this would mean CoRR-T-K, C0RR- S+X, CORR-T+-*X,CoRR-N*-L, CoRR-T->L would not be able to enforce agreement on pairs of segments below the C-V threshold. As pointed out in chapter 4, this would result in unwanted iterative effects, where entire sequences of consonants (or vowels) would be affected by an agreement constraint. Opening up correspondence to all segment types would then have disastrous consequences for the analysis of laryngeal agreement in chapter 4, as well as the analysis of uvularity agreement in the following section.

5.4. Analysis of Lexical Patterns

In order to model the gradient dorsal agreement pattern, correspondence constraints between consonants in a string are adopted (Hansson 2001, Rose & Walker 2004), as was the case in chapter 2. Under the Agreement by Correspondence model (Rose & Walker 2004), similarity is the source of correspondence, and correspondence relations can obtain between

138 segments within strings. The similarity based correspondence hierarchy is reprinted here, for convenience:

(41) Similarity-based correspondence hierarchy CoRR-T4-K, CoRR-S->X>> CoRR-T->X >> CORR-N*->L>>CORR-T-*L

The relevant portion of the hierarchy here is the obstruents. As illustrated in the lexical patterns in chapter 2 and in section 5.2, it is the obstruents (and in particular, the dorsal obstruents) that participate in agreement. With this in mind, the constraint mandating output correspondence can now be developed as one that demands correspondence of two segments, but only if those segments are [-son]. It is worth pointing out again that this constraint is labeled CORR-T-Xto reflect the fact that it is both blind to place of articulation (as in the laryngeal agreement discussed in chapter 4), and also blind to continuancy.

(42) C0RR-T4-->X:Let S be an output string of segments and let X and Y be segments that

are specified for [-son]. If X, Y E 5, then X is in a relation with Y; that is, X and Y are correspondents of one another.

The basis for much of the Agreement by Correspondence framework is the notion of similarity, and a motivating factor behind it is the fact that speech planning seems to be sensitive to similarity. Speech errors are often explained on the basis of similarity, whereby the more similar two speech sounds are, the more chance they will be subject to error (Nooteboom 1967, MacKay 1970, Fromicin 1971, Shattuck-Hufnagel and Klatt 1979, van den Broecke & Goldstein 1980, Stemberger 1991). This has been taken as evidence for similarity as a salient phonological notion (Frisch 1996, Hansson 2001, Walker et al. 2002, Rose & Walker 2004, Rose & King 2007). As somewhat anecdotal evidence for velars and uvulars behaving as similar consonants, Rigsby (1970:212-213) reports a tongue-twister that plays on the phonetic difficulties that “arise from the proximal articulatory positions of its palatals, velars, labiovelars, and 46uvulars.”

46 Rigsby draws phonetic distinctionsbetween low vowels that are not expressed in this thesis.

139 ______

(43)na-qaks-ti ka?-+ {aqaxqa:kkw+ +a-qax”-qak’4 +a-qax-q’a:x-+qa:q [naGáksdi: Gá?a+ +aGaXGá:kx9 +aGctx”Gák”+ 1cIGaq’ á1 Ga:q9 ‘I have just seen for the very first time the toughness of the sinews of the wings of the raven’

Thus, the similarity of velars and uvulars appears to be visible even in sources that might be considered somewhat marginal, or removed from the phonology “proper”. The constraint in (42) ensures that all obstruents are in correspondence. However, like any other OT constraint, it is violable. This is illustrated in (44), where segmentally identical candidates are evaluated based on their satisfaction of CoRR-T-*X. While the candidates in (a-b) and (c-d) are phonetically identical, respectively, candidate (b) and (d), in

which the dorsals are not correspondents, loses because it violates C0RR-T4--*X. (44) illustrates the point that C0RR-T÷-*Xselects dorsal candidates that are in correspondence:

(44) Effects of C0RR-T->X CoRR-T-*X a. k-q b.k-q cc k-x d.k-x

It is worth noting that both agreeing and non-agreeing pairs of dorsals can be correspondents; this points to the crucial theoretical difference between the notion of correspondence and the notion of agreement. With correspondence established by the constraint in (42), then another set of constraints is responsible for governing the agreement between correspondents. In order to force correspondents to agree, there must be a constraint mandating identity with respect to some feature. Formally, the relevant constraint is IDENT-CC[FJ:

(45) IDENT-CC[F]: Let c be a segment in the output and 3 be any correspondent segment of a

in the output. If a is [F], then f3is [F].

140 What type of feature IDENT forces correspondents to agree with respect to is of critical importance here. As demonstrated in section 5.3, the feature in question is

[pharyngeal] — a pair of dorsal consonants will either agree with respect to [phar], or they will disagree. For an active consonantal agreement with alternations to obtain, both C0RR- T->X and IDENT-CC[phar]must outrank either IDENT-IO[pharjor IDENT-OI[phar]. This ranking enforces both consonantal correspondence, and also agreement between those correspondents.

(46) Forcing agreement /kaq/ CORR-T+->X IDENT-CC[phar] IDENT-IO[phar] IDENT-OI[phar] a. kaq b. kak * c.kaq *! d.qaq *

In the tableau above, candidate (a) fails by virtue of its dorsal consonants not being in a correspondence relation. Candidate (c) is likewise eliminated: despite the dorsals being in correspondence, they do not share the same place of articulation (velar vs. uvular). Candidates (b) and (d), with correspondence between output dorsals, and the same place of articulation, are the winners. So far the dorsal agreement phenomenon has been treated as if it was categorical. As was noted above and in chapter 2, however, this effect is gradient in nature. That being the case, the configuration above needs to be modified in order to account for the gradience of the effect. As was demonstrated in chapter 3, the relevant modifications involve merely adopting a set of numerically weighted (rather than ranked) constraints, as in Harmonic Grammar. Under the weighted-constraints approach, outputs will match inputs due to highly weighted faithfulness constraints, with (nearly) all markedness constraints being weighted lower. The learning algorithm was again implemented in Praat, with the same settings as in chapter 4. The weights that were derived from running the algorithm are presented in (47):

141 (47) Constraint weights after learning Constraint Weight after learning IDENT-CC[phar] 100 CORR-N->L 100 IDENT-OI[phar] 96.4 IDENT-IO[phar] 96.3 CoRR-S-*X 81.2 CoRR-T-*K 63.1 CORR-TE->X 7.3 CoRR-T-*L 7.3

As with the constraints on laryngeal agreement in Chapter 4, IDENT-CC[phar]is at the top of the hierarchy with an unchanged weight of 100. C0RR-N+->Lis also at 100, simply because there are no pairs of dorsal sonorants to violate this constraint. The input-output faithfulness constraints IDENT-OI[phar]and IDENT-OI[phar]are near 100, as they ensure output foms surface faithfully. Below these are the bloc of CC correspondence constraints, which started at 100, and were nudged down below faithfulness. These constraints can now be seen at work with cases of agreement (48) and non- agreement (49).

(48) (Gradient) agreement Weight 100 96.4 96.3 7.3 H A /qe:/ IDENT IDENT IDENT CORR ‘to slide’ CC[phar] OI[phar] IO[pharj T->X Ge( 0 103.6 oe:x -1 -1 -103.6 -103.6

Weight 100 96.4 96.3 7.3 H A /kuxs/ IDENT- IDENT- IDENT- C0RR- ‘wake up’ CC[PHAR] OI[PHAR] IO[PHAR] T->X guxs 0 103.7 guys -1 -1 -103.7 -103.7

These input-output pairs above are uninteresting, as the winning candidate does not violate any of the constraints. The result is what can be considered “accidental” agreement, where the input pair already agrees for [PHAR]. Instead, observe cases of non-agreement. Pairs of non-agreeing consonants, such as dorsal-uvular pairs, are illustrated in (52) below:

142 (49) Non-agreement Weight 100 96.4 96.3 63.1 7.3 H A /qa:k’’/ IDENT IDENT IDENT CORR CORR ‘sinew’ CC[PHAR] OI[PHARI IO[PHAR] T-K T+>X kak” -1 -96.4 -26 coa:k” -1 -1 -70.4 25.9 Ga:q -1 -96.3 25.9

Because of the highly weighted faithfulness constraints, the output will consistently surface identical to the input; any candidates that violate these constraints will incur a harmony score less than a candidate that does not violate them (the optimal candidate). In addition, it was argued in Chapter 4 that non-agreeing output forms violate output correspondence constraints (here CoRR-T->X), and not the agreement constraint IDENT-CC[phar]. In other words, non- agreeing outputs contain segments that are not in correspondence, not correspondents that do not agree. Highly weighted IDENT-CC[phar]ensures this state of affairs. Furthermore, if the acceptability scores are compared across the tableaux in (48) and (49), it can be seen how this type of system reflects the structure of the lexicon. Boiled down, this is what it means to have (gradient) pharyngeal agreement: [k-qJ (from underlying /k-q/) is worse than [-q] (from underlying /x-q/)or {k-xJ (from underlying /k-x/). This is the gradient pattern that was discussed in chapter 2 and in section 5.2 of this chapter: velar uvular pairs are more underrepresented than velar-velar or uvular-uvular pairs (which are in fact overrepresented).

(50) Acceptability and OlE for dorsal pairs Dorsal pair - Acceptability OlE uvular-velar 25.9 0.57 velar-velar 103.7 1.40 uvular-uvular 103.6 1.46

It is worth noting here that so far the analysis has assumed that consonantal agreement is driven by a highly weighted IDENT-CC[phar],and a lower weighted CORR

143 T—>X. This weighting, in effect, means agreement is always enforced among correspondents, but that a correspondence relation is not always 47established This situation, however, could in principle be reversed, with much the same result. In other words, it could easily be the case that agreement between correspondents is always. enforced, but that the establishment of a correspondence relation is occasionally violated. Hansson (2007b) identifies this problem, and notes that while blocking effects have not been found in long- distance consonant agreement systems, these effects are predicted to be possible under the former condition mentioned above. Since it was shown in the previous chapter that IDENT CC constraints had to be weighted higher, and C0RR-CC constraints had to be weighted lower in order to attain the proper results during the learning procedure, this will also be assumed for the phenomenon at hand. In the previous chapters, the crucial question of how a grammar with weighted constraints captures (or reflects) the gradient patterns that are found across the lexicon in

Gitksan was raised. It was claimed in chapters 3 and 4 that the gradual learning algorithm for HG “shapes” the grammar each step of the way, until the lexical patterns are molded into the phonology. It is interesting to place the uvularity agreement of Gitksan into a typological perspective. Since it is claimed that it is the phonology of Gitksan that is responsible for the fact that disagreeing pairs of dorsals are underrepresented (i.e. there are fewer than would be expected by random chance), then it must be the case that the mechanism responsible for cross-linguistic variation, the constraint weights, must somehow be different in Gitksan than in other languages where there are no such distributional asymmetries. The difference lies in the weights for the C0RR-CC constraints, which allow for variation along the similarity parameter, and also in the weights for the input-output faithfulness constraints relative to the constraints that impose agreement between correspondents. In a language with categorical agreement, such as Misantla Totonac or Tlachichilco Tepehua, the constraint imposing agreement, IDENT-CC, will be weighted higher than input-output faithfulness (IDENT-lO and/or IDENT-Ol) just like in Gitksan, but with the relevant C0RR-CC constraint also highly weighted. In Gitksan, the existence of a gradient effect is due to the lower weighting of

Having a correspondence relation always established between any homo- or heterorganic pairs of obstruents could prove problematic for laryngeal-velar cases. The prediction here is that a [phar] agreement between these consonant types should arise, with a statistical underrepresentation between velar-laryngeal pairs. However, if [h] and [2] are considered laryngeal sonorants, as is the case in many languages, then this problem evaporates (since obstruent-sonorantpairs do not have correspondenceimposed on them).

144 CoRR-TE->X,which doesn’t play a role in creating alternations, but does play a role in contributing to harmony scores (and thus acceptability scores) of competing candidates. In Gitksan the learning algorithm nudges this constraint down further and further each time a disagreeing pair is encountered, while at the same time nudging IDENT-IO[phar]and IDENT OI[j,harj up. In categorical patterns, on the other hand, this constraint will remain highly weighted, as it will encounter few (if any) forms that violate it.

5.5. Conclusion In this chapter the ‘guttural’ patterns of Gitksan were explored. Drawing on the results from chapter 2, it was shown that there exists a general prohibition on dorsal-dorsal consonant sequences; however when two dorsals do cooccur, there is a pressure for them to share the same place of articulation. Thus, velars tend to cooccur with velars, and uvulars tend to cooccur with uvulars. It was further shown that the uvulars and laryngeals patterned together in two ways: in patterning gradiently over the lexicon as a guttural class that was underrepresented in its cooccurrence, and in the categorical process of vowel lowering. The representation for uvulars was enhanced so that they are specified as [pharyngeal], a shared characteristic of the laryngeals. In order to account for this uvularity agreement, the

Harmonic Grammar approach of chapter 3 was employed. It was then claimed that the nature of the gradience of this pattern was similar to the patterns discussed in chapter 2, and required the same analytical treatment.

145 Chapter 6: Reduplicative Patterns

6.1. Introduction

Having established in chapter 2 that the phonotactic patterns in the Gitksan lexicon reflect both OCP and assimilatory effects, it is now relevant to investigate the phonological patterns across morpheme boundaries. It has been observed by Dunn (1979a) that there are gradient assimilatory and dissimilatory effects in reduplicative contexts in Coast Tsimshian. This being the case, reduplication in Gitksan will be investigated to determine whether the same gradient patterns exist. There is variability in reduplicative template selection in these languages; thus, this chapter will attempt to determine whether phonological environment is a determining factor in template selection. Gitksan also exhibits categorical patterns affecting laryngeal features in reduplicants. When glottalized consonants are copied from the Base to the Reduplicant, they become deglottalized. Placed into a larger Tsimshianic perspective, it is claimed that these patterns are the result of positional faithfulness constraints interacting with markedness constraints. The result of the analysis is that glottalized sonorants are more marked than glottalized obstruents (ejectives).

Section 2 provides a background to the reduplication facts of the language. Section 3 outlines the gradient patterns found in Coast Tsimshian, and presents evidence suggesting that these patterns do not exist in Gitksan. Section 4 outlines the phenomenon of reduplicative deglottalization, and illustrates the differences in how deglottalization behaves among the Tsimshianic languages. Section 5 concludes.

6.2. Reduplication Basics

In this section, the morphology associated with plurals will be investigated. There are several different ways of marking plurality in Gitksan. These include various types of reduplication, prefixation (with l- or distributive Ga-), irregular morphology (including gil

146 mutation, vowel ablaut, and other irregular forms), and 48suppletion While the variation in what types of morphological forms are used by speakers is itself puzzling, a more focused question concerns the reduplicative subset of these options.. This facet of the morphology is interesting because there seem to be no semantic features which guide reduplicant shape (as opposed to the a- prefix, which marks distributive plurality), and also because there is afready an established literature that discusses the issue in Coast Tsimshian (Dunn 1979a). There are three basic reduplicative allomorphs that are made use of by the language: CV-, CVC-, and CVx-. These are illustrated with examples in (1) below. It is important to note at this point the various phonological processes that affect reduplicants. For instance, a prominent feature is that obstruents and sonorants deglottalize in reduplicative contexts; the prevocalic voicing of plosives (discussed in chapter 4) is predictable, and is fed by this deglottalization (deglottalization will be dealt with in section 6.4). There are other leniting processes that affect the coda of reduplicants, such as deaffrication and stop spirantization.

(1) CV- reduplication ts’ak’ cki-ts’ak’ ‘dish’ do?o di-do?o ‘cheek’ baPa bi-ba’?a ‘thigh’ lit li-lit ‘wedge’ gin gi-gin ‘to give’ gida gi-gida ‘to ask’ dalq di-dalq ‘to talk with’

(2) CVC- reduplication ?isx’’ ?as-?isx”’ ‘stink, smell’ chap &ip-cap ‘make, do’ jim jim-jim ‘smell (vt)’ t’eel’t dil-t’eel’t ‘be fast, quick’ Gats Gas-Gats ‘pour’ dulpx” dil-dulpx” ‘to be short’ t’oq da-toq ‘to grab’

Rigsby (1986) also listspleonastic plurals, which consist of the conjunctive use of two or more strategies listed above.

147 masx” mismasx\v ‘to be red (ochre-coloured)’

(3) CVx- reduplication gwej g’ix-g”e:j’ ‘be poor’ huk”sx” haxhukvsxw ‘be there (w/someone), join’ k’e:Gan gix-k’e:Gan ‘drill (vt)’ loGom wil boom wix-wil ‘be old and useless’ bil’ust bix-bil’ust ‘star’ didij’ dix-didij’ ‘to look after’

Simplifying processes affecting reduplication include vowel rounding and lowering, deglottalization, deaffrication, depalatalization, and dorsal spirantization. Vowel rounding and lowering were discussed in chapter 5, and deglottalization is treated at length in section 6.4 of this chapter. Deaffrication, depalatalization, and dorsal spirantization are illustrated below. Deaffrication (4) renders affricates in the base as homorganic fricatives in the coda of a reduplicant.

(4) Deaffrication m’ats mis-m’ats ‘to hit, strike’ t’u:ts’x” dist’uts’x” ‘be black’ (vi)’, ‘knife, metal’

In some 49sub-dialects of Gitksan, palato-alveolar fricatives and affricates exist. When a base containing these fricatives is reduplicated, the fricative surfaces as alveolar:

(5) Depalatalization ma$xw mis-ma$x” ‘white?’ i$xw as-j$x” ‘stink, smell’ gwis..gwo$xw ‘blue’

The term “sub-dialects” is used here, as the phenomenon is not attributable to the Eastern/Western dialect split, nor is it exhibited only in idiolects.

148 Finally, dorsal stops in the coda of a reduplicant spirantize, as illustrated in (6). This phenomenon will be discussed further in section 6.3.1.

(6) Dorsal Spirantization tsi:kv ‘to leak’ tjxw dix/tixw ‘to be big, stout (of a person)’ ts’ak’ dzix-ts’ak’ ‘dish (of several kinds)’

As can be seen from the examples above, the CV- reduplicant type copies the first consonant (Ci) of the base and supplies a vowel of predictable quality, namely [i] (and [a] adjacent to gutturals). The CVC- type copies both the first consonant )1(C and the second consonant )2(C of the base, and provides the same high front vowel [i]. The CVx- type copies the first consonant (Ci), and inserts a fixed velar fricative into the second consonantal slot. It should be noted that all reduplication in the language is almost entirely 50prefixing The relationship between copied portion and base for each allomorph is schematized. in (7). (7) Schematic of reduplicative structure

a. CV- b. CVC- c. CVx i-ts’a k’ d i l-t’e:l’ t d i x-di d i j’ 11111 1111111 1CV 1-C V 2C -CVC 23C 1CVx -CVC tt 12CVt___tj 123

6.3. Gradient Reduplicative Patterns

What is problematic is the fact that it is seemingly impossible to predict which allomorph will be selected with which stem. The situation becomes increasingly complex as one discovers that for a very limited set of lexical items, different speakers may use different strategies to pluralize a stem, or that a single speaker will also offer a plural form using different strategies with little to no effect on grammaticality. It has even been noted that the

° Thereisa very smallclassofinfixing reduplicants,discussedinRigsby(1986).

149 different languages in the Tsimshianic family use entirely different “default” forms of reduplication. As noted by Dunn, “the reduplication data are little short of chaotic” (1979a:59). The most plausible account of the variation in reduplicative allomorphy has been Dunn’s (1979a) treatment of Coast Tsimshian. According to Dunn, while the pattern is seemingly random, the membership of a given form in an inflectional class is to some extent phonologically conditioned, and this phonological conditioning is sometimes variable. Based on Dunn’s conclusions, the hypothesis of this chapter is that the variability in reduplicative allomorphy is based on speakers’ statistical intuitions about phonological environment. Specifically, it can be hypothesized that a speaker of Gitksan is at some level aware of the phonological shape of the stemlbase, and that this knowledge is guiding the speaker in the selection of an allomorph. This is parallel to claims regarding gradient OCP effects in other languages, such as Arabic, where the statistical distribution of consonants in roots is attributed to a speaker’s internal knowledge of consonantal similarity (Frisch & Zawaydeh 2001). In short, following suggestions in Albright et al.’s (2000) analysis of Spanish diphthongization, as a learner of the language, the child combs through the data that is presented to them, looking for phonological generalizations. These generalizations are formed, and the child’s (stochastic) intuitions last into adulthood. It is perhaps important to compare this hypothesis with others at this point. First, since we are dealing with the phonological shape of reduplicants, this is both an issue of phonological regularity, and also one of morphological regularity (see Bybee 1995 for a detailed discussion of morphological regularity and the lexicon). Another important comparison lies in the phonological environment which triggers the allomorphy; in studies such as Albright et al. (2000), it is the adjacent phonological context which helped determine whether a phonological process would apply; in this particular case, we are dealing with potentially adjacent and also potentially non-adjacent phonological environments (for cases of non-local effects, see Aibright & Hayes 2006). Finally, we may contrast the findings of this chapter, where patterns across morpheme boundaries are outlined, with chapter 2, where root-internal tendencies were explored. It will be demonstrated that while the Coast Tsimshian pattern is robust and indeed sensitive to phonological environment, the same is not true for Gitksan. In fact, every tested

150 generalization, categorical or gradient, found in Coast Tsimshian reduplication fails to hold in Gitksan. This being the case, it is hypothesized that the strategy for allomorph selection in Gitksan is fundamentally different than in Coast Tsimshian.

6.3.1. Frequencies of Plural Morphology in Tsimshianic

According to Dunn (1979a), there are five major inflection classes that are expressed as different reduplicant templates in Coast Tsimshian. These include CVk-, CV-, CVC-, CV- and -CV-. These classes differ from those of Gitksan in that (1) there are no forms with a fixed velar stop /k/in Gitksan, (2) Coast Tsimshian lacks a velar fricative, and thus uses the uvular fricative as the fixed segment in CV- forms (see Dunn 1970 for discussion of the underlying status of the velar fricative in Coast Tsimshian), and (3) the CV- reduplicant is not as productive in Gitksan (while in terms of percentages, it is slightly more represented than the CV- form in Coast Tsimshian). The frequency of occurrence of these allomorphs in the Coast Tsimshian lexicon (from Dunn 1979a) is listed in (8) below.

(8) Frequencies of Coast Tsimshian reduplicative allomorphs CVk- 85/183=46.4% CV- 11/183=6% CVC- 49/183 = 26.8% CV- 26/183 = 14.2% -CV- 12/183 = 6.6%

Turning now to Gitksan, we can see whether the same general trends can be found. The frequencies for each of the reduplicative allomorphs in Gitksan is given in (6) below. The total number of reduplicative forms that are unambiguously one of the three allomorphs in the database is 139’.

51 This is a relatively small number, even when compared to the total number of forms collected in Dunn (1979a). A goal for future research is to collect more pluralized and reduplicated forms that will serve as a more concrete basis for comparison.

151 (9) Frequencies of Gitksan reduplicative allomorphs from the lexical database CV- 24/139=17.3% CVC- 97/139 = 69.8% CVx- 14/139=10.1% CV- 4/139=2.9%

A potential problem lies with the indeterminacy between a CVx- and a CVC- form with a

lenited velar stop. As Rigsby (1986) points out, coronal affricates lenite to [s] (deaffrication,

mentioned in section 6.2), and dorsal stops lenite to [x], [x”J, and [yJin 2C position of a CVC reduplicant (dorsal spirantization, section 6.2). This mirrors the general process of lenition affecting some stops in coda position in the language. The problem with these forms is that it is analytically impossible to tell whether a form with a velar stop in the base that seemingly corresponds to 2C of the reduplicant should count as a CVC- or CVx- type of allomorph. In

other words, it’s impossible to tell whether the reduplicant in [CVx-CVk... •]is really a CVC- with a lenited copy of the velar stop, or simply a CVx- with a fixed velar fricative. (10) presents many forms which are ambiguously CVC- or CVx-.

(10) Ambiguous forms (CVC- or CVx-) xii hax-haks ‘to scold, ball out (p1)’ ts’ix-is’ik’ ‘wagon, baby carriage’ six-sak’ ‘to be stretched (p1)’ dix-t+’o:k’ ‘be muddy (p1)’ dix-t’ak’in ‘fold (p1)’

The only disambiguating factor is found in the form of preservation of labialisation; forms that preserve this secondary feature can be considered CVC- forms, while forms which do not appear to preserve this feature can be considered of the fixed CVx- variety. (11a) shows how labialization is generally retained in reduplication, as illustrated by onsets

(compare with glottalization, which is lost in these environments). (1lb) shows how labialization is likewise retained in 2C (coda) position, which distinguishes CVC- from CVx reduplicants.

152 (11) Forms that retain labialization (unambiguously CVC-) a. gwau gwisgwats ‘rusty brown’ gwa1gwa gwjlgwa1gw ‘to be thirsty’ gwitkwow ‘to be lost, gone’ gwa&w gvi1gwallC ‘to be dry’ gwe:j g”ix-g’’e:j’ ‘to be poor’

b. ts’i:k” ‘to leak’ t’ix’ dix’’-t’ix” ‘to be big, stout (of a person)’ guxv gux’’-gux” ‘to shoot (trans)’

Because of this indeterminacy, forms like those in (10), which numbered 7, were not included in the frequency breakdowns in (9). For illustrative purposes, the frequencies of forms for the two languages are compared in (12):

(12) Frequency of occurrence of reduplicative allomorphs in Coast Tsimshian and Gitksan Coast Tsimshian Gitksan CV- 14.2% 17.3% CVC- 26.8 % 69.8 % CVx- n/a 10.1% CVk- 46.4 % n/a CV- 6% 2.9% -CV- 6.6% n/a

It is obvious from these numbers that not all reduplicative allomorphs are treated equally across the two languages. For instance, Dunn (1979a) notes that CVk- is the default in Coast Tsimshian, and Rigsby (1986) notes that CVC- is the default reduplicant in Gitksan. Rigsby (1986) goes on to note that the default is not fixed, and that different speakers may rely on different default strategies. I have noted this same trend in my own fieldwork on the language. Returning to the issue of CVC- vs. CVk- defaults, if the CVk- reduplicant has diachronically changed to CVC- (or if the CVC- reduplicant underwent a diachronic split), this would give us roughly the same numbers for Gitksan and Coast Tsimshian. That is, if

153 CVC- and CVk- are summed for each language, the results are 73.2% for Coast Tsimshian, and 69.8% for Gitksan, which are comparable percentages. This comparison lends support to the idea that the CVC- and CVk- classes in both languages are perhaps diachronically 52related.

6.3.2. Reduplicative Allomorphy in Coast Tsimshian

An outline of the variability of reduplicative forms in Coast Tsimshian has been provided by Dunn (1979a). In this work, Dunn outlines the various allomorphs, lines them up in terms of productivity and use, and investigates the phonological and morphological conditioning for each of the allomorphs. Dunn (1979a) is unique among analyses of reduplication within the language family in that he notes the role that phonological similarity plays in the selection of allomorphs. He also notes that this is sometimes a statistical tendency, with an overall effect that is gradient in nature. Dunn (1979a) notes that there are both categorical and gradient determinants of reduplicant shape. For the categorical effects, he notes:

i) when 1C in the base is a 53non-coronal and part of a consonant cluster, then the CV- prefix will be selected. ii) when 1C of the base is 54coronal and part of a cluster, then either the CVk- or CVX- prefix is selected (i.e. CV- or CVC- are not selected).

iii) No CVC- class words have base-initial clusters

iv) Every word that takes CV- has a uvular in the word (not necessarily in the portion that is copied) ex: be?, bci-be? ‘tear up’, ?a:dzoq, ?-?adzq ‘reach across’

52 But this argument holds only if the 97 CVC- bases in Gitksan are cognate with members of the 49 CVC- and 85 CVk- bases of Coast Tsimshian. At present, these bases in the two languages have not been compared. Dunn (1979) notes specifically /p, k, x/. Dunn (1979) notes specifically /1, n, sI.

154 For the gradient effects, Dunn performed a x2 test for non-random association to determine if there are any statistically significant tendencies relating to allomorph selection and phonological environment. The results of this test are gradient, and indicate that:

i) place of articulation of 1C plays a role in CVk- forms (i.e. place is not a random factor; x2 = 20.05, p < 0.001). CVk- forms tend to have an alveolar consonant as 1C of the base 2(x = 8.99, p < 0.01). 60% of CVk- forms have an alveolar as 1C of the base, and 59% of alveolars that are Ci in the base take CVk- as a reduplicant. As a consequence of this, CVk- forms tend not to have dorsals as i in the base 2(X=13 .5, p <0.001). 10% of CVk- words had dorsals as 1C in the base, and 21% of dorsal initial words took CVk- reduplicants.

ii) place of articulation plays a role in CVC- forms (i.e. place is not a random factor; x2 =

19.8, <0.001). CVC- forms tend to have the p dorsals in 1 of base 2(x 15.5, p < 0.001). 45% of CVC- forms have dorsals in 1C of the base, and 51% of dorsals take CVC- reduplicants. Consequently, alveolar base-initials are dispreferred with CVC forms 2(x = 10.7, p < 0.01). 27% of CVC- forms have bases with an alveolar in Ci position, while 15% of alveolars take CVC- reduplicants.

Dunn further notes that 2C of the base is probably a random factor in allomorph selection. Overall, Dunn’s findings are that there is a tendency towards dissimilarity in the Cs of CVC- 55forms (which he terms the “preferred syllable type”), where onset and coda are dissimilar. An alternative way of looking at this is to consider the dissimilarity between 2C of the reduplicant and 1C of the base, such that there is an alternation in place of articulation when moving from the reduplicant into the base. In other words, there is a possible relation between the dissimilarity requirement on CVk- reduplication and the general statistical tendency to avoid similarity in CVC cases. The CVk- and CVC- classes were the primary focus of study in Dunn (1979a), simply because they constitute the largest class, they are the most regular, require the simplest morphological strategies, and are the most productive reduplicative forms. This is schematized in (13) below, where size, regularity, simplicity, and productivity of the class is

Where “CVC-” refers to an abstraction over {CVC-, CVk-, CV-}.

155 laid out horizontally, with each allomorph scaled along this continuum (from Dunn

1979a:62)

(13) Size, regularity, and productivity of Coast Tsimshian reduplicative allomorphs large class, small class, regular, irregular, simple morphology, complex morphology, productive not productive

CVk (CV-) CVC- CV- -CV-

6.3.3. Reduplicative Allomorphy in Gitksan

The Coast Tsimshian data presented above, especially the research by Dunn (1979a), provides a useful backdrop to test hypotheses about reduplication in Gitksan. We can first start out by laying out any categorical effects that can be found, and then address any gradient patterns in the data. One of Dunn’s gradient generalizations about CVC- forms is that the coda tends to be different in place of articulation from the onset. While there are matching dorsal onsets/codas in Gitksan reduplicants, there are also several examples of heterorganic onsets/codas:

(14) non-coronal-Cl bases select CVC a. gux’’ gux’’-gux’’ ‘to shoot (trans)’ b. esx” Gas-Gesx’’ ‘be slender’ c. basiGan bis-basiGan ‘split up, separate (trans)’

Furthermore, Dunn’s generalization concerning bases with initial consonant clusters in Coast Tsimshian also doesn’t always hold in Gitksan, as is evidenced by the form below. In Coast Tsimshian, CVC- reduplication never takes place with consonant cluster initial bases; in Gitksan, this type of reduplication is possible:

156 (15) tq’a da-tq’a ‘skin, hide’

Dunn’s generalization that in Coast Tsimshian it is the coronal-initial bases that tend to have CVk- reduplicants is true (though the numbers are small), though the generalization that dorsal-initial bases tend to have CVC- reduplicants does not hold in Gitksan, since if the coronal- and dorsal-initial bases (within CVC- and CVx-) are compared, there are more coronal-initial bases for both allomorphs.

(16) CVC- and CVx- base-initial consonants Reduplicant coronal-initial base dorsal-initial base CVC- 38 23 CVx- 7 3

The categorical generalization that CV- forms in Coast Tsimshian have a uvular somewhere in the word is also not true for Gitksan, simply because of the fact that Gitksan has no documented regular CV- reduplicant. For instance, of the 13 CVx- allomorphs in the database, 6 have velars. It may not be the case that the uvular fixed segmentism in Coast Tsimshian correlates directly with the velar fixed segmentism in Gitksan; however, there exists sporadic cases with uvulars, and these were also checked. Of these, 3 occurred with uvular-containing words. Again, this is definitely not a categorical generalization of Gitksan, although it might qualify as a gradient one, as 6 velars and 3 uvulars equal 9 cases of words with a dorsal in them and a CVx- reduplicant (out of 13 total cases of CVx- reduplicants).

(17) CVx- plurals max-maGaj’ rainbow mix-muk” purple (p1) gix-ga?a to see, look at dix-didij’ to look after (p1) +ix-+isin to finish (p1) ljx-+isx”' to finish (p1) gwixgwe:j be poor (p1) bix-bil’ust ‘star (p1)’

157 six-saksx” to be clean (pi) six-se:Gal to be rough (collective p1) hax-huk”sx” be there (with someone), join (p1) six-scian’ist mountain (p1) to fall down (p1)

Overall, none of the generalizations that have been made for Coast Tsimshian reduplicants by Dunn are found in Gitksan. To confirm this, the total list of bases and the total list of matching reduplicants were compared. The result of a logistic regression analysis shows that the template shape cannot be significantly predicted from the place of articulation of the base initial consonants (Wald Stats, x2 = 2.59, d.f. 4, p=O.629). Since none of the generalizations that hold for Coast Tsimshian hold for Gitksan, the question arises as to what the constraints are, or what exactly the generalizations are that govern allomorph selection in Gitksan. Since phonological environment does not obviously play a role in allomorph selection, then the child learning the reduplicative system does not have a heuristic to guide them in composing rules or generalizations. That being the case, it is likely that the system must be learned lexical item-by-lexical item. In order for the grammar to correctly choose the proper reduplicative allomorph, any given lexical item must be indexed to a particular templatic constraint. In stepping back to generalize, it can be stated that the shape of the reduplicants must be derived by a set of templatic constraints, such as RED=CVC,etc. These are the relevant constraints under Classical Template Theory of (Marantz (1982), McCarthy & Prince (1986)). Under this approach, the shape of the template is encoded directly into the substance of constraints. As with other phenomena in Optimality Theory, it would be desirable to achieve the same results through constraint interaction alone, and not through stipulated templatic constraints. One way to achieve this is by deriving the template shape from other properties of the grammar, an approach which ultimately views template shapes as epiphenomenal. This approach is termed Generalized Template Theory (McCarthy & Prince 1999). While this approach correctly derives template shapes in other languages by means of constraints on prosodic weilformedness, etc., it does not work for the Gitksan facts. This is because there are no other phonological effects in the language which could conspire to derive the correct reduplicative allomorphs. For instance, CV-, CVC-, and CVx

158 reduplicants all occur with monosyllabic, bisyllabic, and trisyllabic words, so the selection of allomorph is not the result of satisfring a demand on foot weilformedness. This being the case, generalized template theory must be rejected, and the template shapes must be encoded directly into constraints for this grammar to derive the correct outputs. Following Shaw’s (1992) analysis of Nisgha, it will be assumed that reduplication copies enough material for a syllable, but that the template for the syllable includes an onset, a nuclear mora, and a (moraic or non-moraic) coda. In addition, there must be a template that allows for only a CV- reduplicant; these template shapes are not derivable by the interaction of constraints. This being the case, templatic reference to the skeleton must be made in order to determine these reduplicant shapes (Marantz 1982). Such constraints would be formulated as in (18-19):

(18) RED= CVC: Reduplicants are CVC

(19) RED = CV: Reduplicants are CV

Drawing on work by Alderete et al. (1999), it will be assumed that since the fixed segmentism in Gitksan involves the velar fricative [xj, and that this segment is not considered one of the more unmarked segments, then the fixed segment must be specified along with the reduplicant. In order to best illustrate this, there will be a separate constraint REDCVX:

(20) RED = CVx: Reduplicants are CVx

Next, the actual mechanism of indexing templates to lexical forms must be addressed. Lexical indexation does little more here than models a subcategorization relationship between a lexical word and a reduplicative allomorph. Each templatic constraint is given an index (1,2,3), and this index is also found on subsets of lexical items (the ones which subcategorize for any given allomorph). Such subsets would be organized in the lexicon based on the template that each subcategorizes for:

159 ______

(21) 1LEX {Ikat/ ‘man’, /ts’aq’/ ‘dish’, /t’aXI ‘lake’, /qo:t/ ‘heart’, I+ak/ ‘crooked’, etc.} 2LEX {/+ap/ ‘deep’, /kux’7 ‘to shoot’, /t’is/ ‘large, big’, /k”alk”aX/ ‘thirsty’, etc.} 3LEX {/lKwe:jI‘be poor’, Ipil’ustl ‘star’, /ka’?I‘see, look at’, Imuk”/ ‘purple’, etc.}

In order to prevent complete copying of the base, any given templatic constraint must outrank MAX-BR. The analysis here is considering the fixed [i] in the reduplicant to be a (unfaithful) copy of the base vowel, so having a missing vowel in the reduplicant would incur a violation of MAX-BR.

(22) MAX-BRin longer strings /ba+)an! ‘to spread out’ RED=CVC MAX-BR b-ba+an *! ***** bi-ba+an **** bi+-baian *** bi1-ba1anJ * * bi4-a-ba1an *! * bi1an-ba1an

With these constraints in hand, there can be two strata of constraints, where the set of templatic constraints outranks Max-BR: {REDCV, REDCVC, REDCVx} >> MAX-BR. Turning back now to the idea of lexical indexation, (x) illustrates how indexation works for forms subcategorized for CV-, CVC-, and CVx- reduplicants. In each case, violations are only assigned for forms that share the same index as the constraint in question.

(23) Lexical Indexation RED=CV REDCVC REDCVX MAX-BR 1/taw/ di-daw 1 2 3 * diw-daw dix-daw *! * 2/t’is/ di-t’is *! * dis-t’is dix-t’is *! * /muk’”/ mi-muk” *! * 3 *! mixmuklv *

160 It is an issue as to whether all forms must be lexically indexed, or whether one form serves as a default. If there is a default, then only the two other templates must be indexed. Under this scenario, if there is an input with no indexation, the default allomorph will be selected. Since CVC- is considered the default reduplicant in Gitksan, then the ranking is 1{RED=CV }2REDCVX >> RED=CVC>> MAX-BR. (24), The role of default templatic constraint (hypothetical) /kat/ REDCV REDCVX REDCVC MAX-BR ki-kat 1 2 *! * • kit-kat kix-kat *! *

This analysis assumes that if there is no index associated with the lexical item, it will incur violations of any applicable constraints. If a lexical form is indexed, then it is in effect protected from some of the templatic constraints. For a nonce form, there are no indexes at all, so this analysis would predict that all nonce forms would take a CVC- reduplicant, regardless of constraint weightings. The option at this point is to assume that there is a set of indexed templatic constraints as well as a set of non-indexed templatic constraints. Since nonce forms lack an index, it is only the non-indexed constraints that will be applicable, and the relative weightings of these constraints determine the frequency of reduplicant allomorphs applied to nonce forms. This is not a desirable approach, as it requires doubling up on constraints, as well as duplicating constraint weightings for each set of templatic constraints. Rather, a more preferable route to take is to consider all forms in the lexicon as indexed, and that constraints are evaluated by means of whether a lexical form conflicts with a particular constraint indexation. In this way, a lexical item that is indexed as LEXL will incur violations of 2RED=CV and R13r=CVx but will not violate REDCVC as they do not have conflicting 3 LEX, indices. In the same way, which stands for a1nonce word, has no index, and hence will not violate either 2RED=CV REDCVx or RED=CVC, Then it is up to the relative constraint weights, and 3 some, stochastic ranking of those1 constraints, to determine the relative frequency of allomorphs that should surface in the data..

161 6.3.4 Testing Reduplicative Frequencies

In order to test whether the lexical distribution in Gitksan is comparable to that found in novel words, a nonce-probe task (or “wug” test; Berko 1958) was performed. This type of experiment is designed to tap into the productivity of a particular rule or process. In Berko’s original experiment, children were presented with nonce forms, and were asked to provide a plural (as well as other inflectional categories, such as diminutive and past tense) of the item. The reason for using a non-existing form is to test whether the child is applying a productive rule to the form in question, or whether the child is relying on a memorization strategy instead of a rule-based system. In recent years, this type of test has been used to test adult productivity. What this particular test is concerned with is the productivity of each of the reduplicative forms, with the ultimate goal of determining whether there are favored or unfavored environments for each of the allomorphs (Bybee & Moder 1983, Prasada & Pinker 1993, Albright et al. 2000, Albright & Hayes 2006, Hayes & Londe 2006), and whether the lexical frequencies are comparable to frequencies of forms applied to nonce words.

6.3.4.1 Stimuli

In order to create a set of stimuli for the task, a list of nonce words was produced. In order to do this, the list of all possible CVC roots was first generated, then the set of roots that occur in the database was subtracted out from this list. The list was then randomized, and the first 100 nonce roots were selected as stimulus items. A list of existent CVC roots was then selected and was added to this list as a distractor set. This list comprised 50% (n=100) of the entire list (n = 200). All of these roots employed reduplication as their method of pluralization in the database. In order to avoid the potential bias posed by overloading the stimulus set with one particular reduplicant form or another, the percentages of existing CV-, CVC-, and CVx- taking-roots were made to roughly resemble the percentages in the lexicon (see example 6 above). Thus, out of the 100 existing roots, CV- forms consisted of n=21, CVC- n=66, and CVx- n=13. While it was impossible to select an entire set of roots with the shape CVC, as many of these roots were selected as were available, in order to keep the nonce and real roots as prosodically consistent as possible. Since the relative grammaticality

162 of forms was not the subject of this task, only phonotactically well-formed CVC sequences were employed as (nonce) stimuli.

6.3.4.2. Subjects

Two native speakers of Gitksan were tested in this experiment. Both speakers are literate in the language.

6.3.4.3. Procedure

Subjects were presented the example in (25), which used images to guide the subject to distinguish singular from 56plural.

(25) Gya’as Johnhl (os)! ‘John saw a dog’ Gya’as Johnhl (hasos)!! ‘John saw dogs’

Subjects were then instructed to use this same carrier phrase with the list of roots. This was done so that a singular (presumably bare nonce form) would appear in the first instance, and a plural form would appear in the second. Since Gitksan allows plurality to be expressed on both nouns and verbs, with either the meaning of plurality of event or the expression of plural agreement on verbs, the set of verbal roots (in addition to nominal roots) is entirely appropriate. The stimuli were then presented to literate subjects, who provided the plural forms. Subjects were informed that the stimuli consisted of several Gitksan words, though some might be rare, archaic, or loans from neighboring Coast Tsimshian. Subjects were also told that some forms may not be recognizable as Gitksan words, but that they should make their best attempt at producing a plural form.

56 This type of examplewasusedto prime the subjectto derive plural forms. Obviously this sentence frame would not accommodate verbal forms. However, to ensure that the idea that plurals were the target structure (for both nominal and verbal roots), this frame was used, and the first couple of words on the list were nominals. After the first few nonce plurals were produced, the experimenter would check with the subjects to make sure that the task was understood, and subjects were informed that they no longer needed to fit the words into the carrier phrase.

163 6.3.4.4. Results

Since all living Gitksan speakers are bilingual in English, it is first necessary to show that the forms offered by subjects were in fact treated as linguistic forms, and not as artifacts of English, or treated as an English word game. The fact that speakers employed the phonological rules of Gitksan to these forms attests to this. For instance, there was evidence that speakers employed the rule of reduplicant deglottalisation (Brown 2007), coda spirantization, vowel lowering (chapter 5), and also the rule of prevocalic plosive voicing (chapter 4) where relevant:

(26) ts’eq (nonce) dza-ts’eq

Speakers would also regularly de-affricate in the coda position of reduplicant forms, which is another regular rule of Gitksan phonology:

(27) k”its’ (nonce) gwiskwits

Furthermore, the labial cooccurrence restriction (discussed in chapter 5) found in reduplicants was found to be in effect in the nonce forms:

(28) w’op’ (nonce) hu-w’op’

Also, the fact that a range of different strategies was used strongly suggests that speakers were not relying on a “default” strategy of simply doubling the base form; instead, speakers treated each nonce form as a unique item that required a unique analysis by their phonology. As noted above, almost every single type of pluralization strategy was used in the nonce-probe task. We can here see the raw numbers and frequencies of each of these strategies in (29). It is interesting to note that there were several CV- forms (with uvular fricative) for Subject 1, an allomorph that does not seem to be productive in Gitksan. (29)

164 gives the frequency of occurrence of each pluralization type, with data pooled for both subjects (n200).

(29) Frequency of occurrence of pluralization type in nonce-probe task (n = 200) CV- 34.5% (69/200) CVC- 16.5% (33/200) CVx- 5.5% (11/200) CV- 4% (8/200) other 39.5% (79/200)

We can notice individual variation by separating out the results for each subject and breaking the frequencies down using the same categories:

(30) Frequency of occurrence of pluralisation type in nonce-probe task (n = 200)

Pluralization Type Subject 1 = Subject 2 CV- 36%(36/l00) 33%(33/l00) CVC- 6% (6/100) 27% (27/100) CVx- 5%(5/100) 6%(6/100) CV- 0%(0/100) 8%(8/100) other 53%(53/l00) 26%(26/l00)

In order to further refine these frequencies, the class of “other” (non-reduplicative) plurals was removed from counts. Once this is done, then only unambiguous reduplicants are left, and therefore the relative frequency of each reduplicant type (within the set of reduplicants) can be measured. This will allow for the cross-comparison between frequencies found in the Gitksan lexicon with frequencies in the nonce-probe task. The frequency of each reduplicative allomorph, relative to the total number of reduplicative forms volunteered in the nonce-probe test is given in (31) below. When this class is isolated, the average frequencies increase.

165 (31) Frequency of occurrence of reduplicants in nonce-probe task Pluralization Type Subject 1 Subject 2 CV- 76.6% (36/47) 44.6% (33/74) CVC- 12.8%(6/47) 36.5%(27/74) CVx- 10.6%(5/47) 8.l%(6/74) CVI- 0% (0/47) 10.8% (8/74)

These totals can be compared to the frequency of occurrence in the lexicon, as discussed in (9) above. In other words, under the hypothesis that speakers are accessing internalized phonological knowledge in order to provide a plural for the nonce form, the distribution of forms (i.e. which allomorphs are present) should be a statistical reflection of the existing database, rather than a random distribution. By incorporating a similar percentage of each reduplicant type into the experimental design, this will hopefully prevent any kind of bias in terms of accidentally frequent forms, or bias in terms of default strategy. If we compare the two sets of frequencies, we can see that some of the numbers are fairly close, with CV- forms occurring more often than usual in the nonce-probe task (and the CVC- forms occurring much less than expected).

(32) Lexicon vs. wug frequencies Allomorph Lexicon Wug CV- 17.8% 57% CVC- 71.9% 27.3% CVx- 10.4% 9%

(33) LexicorLvs. wug Irequencies (b subject) Allomorph Si S2 Lexicon Wug Lexicon Wug CV- 17.8% 76.6% 17.8% 44.6% CVC- 71.9% 12.8% 71.9% 36.5% CVx- 10.4% 10.6% 10.4% 8.1%

It seems that while frequencies that Subject 1 employed are fairly close to those of the actual lexicon, subject 2 shows more variation. For instance, Subject 2 exhibited a much lower tendency to use CV- forms than subject 1, though a higher tendency in using CVC forms. For CVx- forms, subjects 1 and 2 are under and over the lexical frequency,

166 respectively, but both numbers are considerably low, and roughly the same percentage from

the lexical frequency (10.6% for Subject 1, 8.1% for subject 2). Overall, the percentage of allomorphs used by subject 2 may be skewed due to the smaller number of total reduplicants used (as compared to non-reduplicative forms). For

instance, there were 20 instances of the [-dit] suffix and 19 instances of the [-SIsuffix. It is

assumed that [-ditj is being used as the 3 person plural pronominal suffix, and [-s] is being

used as the English plural. In addition, there are 11 instances of a [-da] suffix, of which I

know no comparable morpheme in Gitksan. The high use of “other” forms for this subject may be forcing the reduplicative numbers to be smaller, and hence skewed. Future experimentation in this area, in which reduplicative vs. non-reduplicative allomorphs are controlled for will be able to shed more light on this difference.

6.3.5. Discussion

It was demonstrated in chapter 2 that within roots, the occurrence of dorsal-dorsal pairs of consonants is significantly underrepresented. However, when the class of dorsals was opened up to investigation, it was found that uvular-uvular and velar-velar pairs were actually overrepresented, and it was claimed that this constituted a kind of gradient consonant agreement. While the actual statistics of cooccurring consonant pairs is not available at

present for the Coast Tsimshian lexicon, the results of Dunn (1979a) are suggestive. What is

interesting to note here is 1) the generalized dissimilatory pattern found between i and 2C of the reduplicant, and 2) the specific assimilatory pattern found in CV- forms, whereby a

uvular is found somewhere in the base word. However, this pattern did NOT hold in Gitksan, a surprising result (section 6.3.3). This being the case, it was determined that similarity does not play a role in Gitksan allomorph selection, and instead, each potential base must be lexically indexed, or subcategorized for a specific template. The hypothesis that the relative frequencies of allomorphs in the lexicon represents lexical distributional knowledge on the part of the speaker, was tested, and results indicate that with respect to CV vs. CVC- templates, the comparison did not yield similar results, but the CVx- vs. CV-/CVC numbers do match quite closely.

167 6.4. Patterns of Deglottalization

The goal of this section is to provide evidence from reduplication in the Tsimshianic languages which shows that glottalized sonorants and ejectives exhibit an asymmetry. The typology of deglottalization processes in the family, most likely due to diachronic changes, illustrates how glottalized sonorants and obstruents are variably targeted. This pattern of variation provides evidence for a scalar view of markedness (Prince & Smolensky 1993 [2004], deLacy 2004), whereby constraints are ordered according to a markedness hierarchy. Furthermore, additional evidence from these languages supports Maddieson’s (1984) implicational hierarchy, whereby glottalized sonorants are more marked than ejectives. All three languages have ejectives and glottalized sonorants in their inventories, and these segments appear in both prevocalic and postvocalic positions. In all three languages, there is a process of deglottalization which neutralizes glottalized consonants in the reduplicant. In 2C position, deglottalization is invariant and affects both obstruents and sonorants. 1C position exhibits more variation. Coast Tsimshian displays no degloftalization in this position; in Nisgha deglottalization affects only sonorants, and in Gitksan it affects both obstruents and sonorants. Before observing the relevant data., it is first necessary to establish some representational assumptions about deglottalization. Following Fallon (2002), this process is assumed to be represented as the delinking of the [constricted glottis] feature from the Laryngeal node:

(34) Rt

Lar 4 [cg]

As Fallon notes, the process of deglottalization is distinct from delaryngealization, which delinks the entire Laryngeal node from the Root node:

168 (35) Rt

Lar

[cg], [sg], [voice]

The piece of evidence which points towards the representation in (34) rather than that in (35) is the maintenance of plosive voicing in reduplicants:

(36) bal’ bil-bal’ ‘feel’ daw di-daw ‘ice, to freeze’ ckam cim-ckam ‘cook, boil’ gat gi-gat ‘to be born, hatch’ gwau g’’is-g”ats ‘rusty brown’ Go:t Ga-Go:t ‘heart’s’

It is important to note that (34) and (35) are representations of pre-OT/HG rules, and not necessarily of output representations. For these cases, especially (35), the constraint OBs-volcE will force the insertion of [voice] on any plosive that is prevocalic. This entails either having or inserting a laryngeal node in the output. Thus, while it will be shown that glottalized consonants can become deglottalized in

certain contexts, there is never a case of devoicing — that is, OBs-vol is weighted higher than any wholesale ban on laryngeal features. Furthermore, when plosives in the reduplicant are deglottalized, they are subject to the process of plosive voicing:

(37) t’is dis-t’is ‘to push’ ts’ak’ dzi-ts’ak’ ‘dishes (of one kind)’ ko:txw ‘to be lost, gone’ q’ots Gas-q’ots ‘to cut’

This being the case, if [voice] is dominated by the Laryngeal node, then in cases of deglottalization, this node must still be present, otherwise the plosives would surface as pulmonic voiceless, in violation of the constraints on voicing.

169 With these representational assumptions in hand, we can now move on to the actual deglottalization data, starting with coda deglottalization and moving on to onset deglottalization.

6.4.1. Coda Deglottalization

In Coast Tsimshian, Nisgha, and Gitksan, deglottalization affects obstruents and sonorants in the 2C position. Ultimately this consonantal slot will always amount to being a syllable coda, as there is a requirement of the base that it have an initial onset. This deglottalization effect is illustrated below for each 57language where obstruents are shown in the (a) examples and sonorants in the (b) examples., (38) Gitksan coda deglottalization a. hit’ hit-hit’ ‘scar, to heal’ hit’ hat-hit’ ‘stick, adhere to’ b. bal’ bil-bal’ ‘feel’ +i-bal’ li-bil-bal’ ‘rub, massage (trans.) haw’ haw-haw’ ‘go home’ (Hunt 1993:161)

(39) Nisgha coda deglottalization a. hit’ hat-hit’ ‘to stick’ tIk’ tix—tik’ ‘to feel silly, shy’ (Shaw 1987:302) b. tam’ tim-tam’ ‘to press sthg’ qIn’ qan-qIn’ ‘(sg) to chew, to chew sthg’ ts’ál’ ts’il—ts’ál’ ‘face, (pair of) eyes’

(40) Coast Tsimshian coda deglottalization a. ?a:p’aq ?ap-?a:p’aq ‘to remember’ sI:p’on sp-sI:p’n ‘to love’ wá:q’ wax-wá:q’ ‘to dig’ qO:jp’a qap-qo:jp’a ‘bright’ qó:jt’ks qat-qO:jt’ks ‘to arrive’ +ájk’a 1k-1ájk’a ‘scar’

Unless otherwise noted, data from Coast Tsimshian is from Sasama (1995) and data from Nisgha is from Tarpent (1983). There is some variation among transcriptions for the different languages, although the differences are not crucial to the points being made in this paper. The reader is referred to the original sources for explanation of transcriptions.

170 hat’axk hat-hat’axk ‘bad, spoiled’

b. k’ám’l k’3m-k’ám’ol ‘to pinch’

In each case, deglottalization can be shown to affect both the obstruent and sonorant series. While the pattern for 2C is invariant across the language family, this can be contrasted with the pattern for Ci, illustrated in the next section. Since there are many languages which have ejectives but not glottalized sonorants (but almost no language with glottalized sonorants and not obstruents; Maddieson 1984), it will be assumed that there are two separate paradigmatic markedness constraints governing each type, *cG/soN and *cG/OBs, the definitions of which are modified from Howe & Pulleyblank (2001):

(41) *CG Specifications of the feature [cg] are prohibited

(42) *CG/SON Specifications of the feature [cg] are prohibited on [+sonorant] segments

(43) *CG/OBS Specifications of the feature [cg] are prohibited on [-sonorant] segments

While there is no evidence at present to force the distinction between glottalized obstruents and sonorants, data from the next section will illustrate that it is necessary. Thus, for the time being both constraints will be collapsed into one: *CG. To this we can add the relevant faithfulness constraints:

(44) IDENT-IO(cg) (ID-JO[cgj) Segments specified for [cg] in the input have identical specifications for [cg] in the output correspondents

(45) IDENT-BR[cg] (ID-BR[cg]) Segments specified for [cg] in the base have identical specifications for [cg] in correspondents in the reduplicant

171 In order for both the ejectives and sonorants to surface in their normal distributions, the faithfulness constraint ID-IO[cg] must dominate both markedness constraints. However, to correctly derive the surface patterns in the reduplicants (in which deglottalization occurs), *cG must dominate ID-BR[cgj.

(46) Ranking deriving coda deglottalization

ID-IO[CG]>> >>ID-BR[cg]

(47) Tableau for Emergence of the Unmarked effect RED+hit’ ID-IO(cg) KCG ID-BR(cg) a. hit-hit’ * * b._bit’-hit’ *! • c. hit-hit

This ranking derives the classic Emergence of the Unmarked type of effect (McCarthy & Prince 1994). While deletion of glottalized segments is banned in the mapping from input to base (enforced by the ranking of ID-IO[cg] 58>>*CG) deletion of glottalized segments in reduplicants is allowed by the domination of ID-BR[cg], by the markedness constraint *cG. In the next section we will see how patterns of onset deglottalization require the separate *cG/soN and *cG/oBs constraints, and how these constraints need to be individually interleaved with faithfulness constraints. Importantly, this section has presented an analysis for the pattern of coda deglottalization, a pattern which seems to be present across all languages of the family. From this point, it now becomes relevant to treat the remaining pattern, that of onset (or )1C deglottalization, within each individual language.

6.4.2. Onset Deglottalization The ,1C or onset pattern can be shown to be a scale with Coast Tsimshian at one extreme (preserving glottalization on all consonant types), to Gitksan at the other (deglottalizing all

This relatively high ranking of IDENT-IO[cg]prevents overapplication or backcopying effects in the base. There is a process of glottal dissimilation in Gitksan that affects sonorants in connected speech contexts over a word boundary (see Rigsby 1986:182) where IDENT-IO[cg]would be violated, though this process will not be elaborated on here.

172 consonant types). In between lies Nisgha. This pattern will be explained for each language, moving from Coast Tsimshian to Gitksan, then addressing the Nisgha data.

6.4.2.1. Coast Tsimshian In Coast Tsimshian, consonants in the onset of a reduplicant are never 59deglottalized This is true of both obstruents (as seen in 48a) as well as sonorants (48b). . (48) Coast Tsimshian retention of glottalization a. p’ó: p’k-p’ó: ‘broken’ t’ó:q t’ax-t’ó:q ‘to suck’ c’O: c’k-c’ó: ‘to skin (animals)’ k”á:n’ kThtkwá:n ‘to lose’ k”’ás ‘to break (trans.)’ (Dunn 1970:52) q’á: q’a-q’á: ‘wound’ b.60w’á: w’t-w’á: ‘to fmd’ (91) w’ó w’u-w’ó ‘quest (noun)’ (Dunn 1970:54) m’ák m’ok-m’dk ‘to catch (on a net)’ (84) j’aq j’á:-jaq ‘to hang’ (95)

In sum, deglottalization does not affect the onset of a reduplicant in Coast Tsimshian, but as seen in (48), codas are deglottalized. Given the original ranking established above, only a slight modification need be made to account for this pattern. A positional faithfulness constraint (Beckman 1997) which mandates faithfulness to the syllable onset of reduplicants must be added to the ranking:

Sasama (1995:87) identifies only 2 words which deglottalize in this position — k“ham,k’ilk“ham ‘to give’ and q ‘ó:lq,qalq ‘ó:lq ‘dull’. Sasamaattributes these exceptions to optionality, where some words are pronounced as ejectives or as plain stops. Sasama even cites a varied pronunciation of the first example as k“ilk“ilám or as k’ilk’ilám. 60 There are some discrepancies in many of the Coast Tsimshian forms available. For instance, there are forms in Dunn (1995) which show both retention and deletion of glottalization on sonorants (where = [ii,<‘y> = [j’J): ‘yuuta yik-’yuuta ‘man’ ‘yuutk ‘yik-’yuutk ‘carry around the neck’ (pg. 17) There appear to be no such exceptions in Sasama’s (1995) data. This may be an indication of idiolectal variation, of sound change in progress, etc.

173 (49) IDENT-BRGONS[cg] (ID-BRG0Ns[cgj) Segments specified for [cg] in the syllable onset of the base have identical specifications for [cg] in correspondents in the reduplicant

Thus, under an OT approach, a constraint such as ID-BRO0Ns[cg]must be undominated, outranking the markedness constraint *cG, as well as the non special variant ID-BR[cg]. For comparative purposes, *CG will now be broken down into its component constraints *cG/soN and *CG/O3S, although there is no evidence at this point in Coast Tsimshian to suggest that they are crucially ranked with respect to each other:

(50) ID-IO[cgj, ID-BRG0Ns[cg] >>*c3/so, *CG/OBS, ID-BR[cgj

However, in HG, constraint violations are additive, meamng ID-BRcYONS[cg]does not have to be highly weighted in order to derive the correct results. Instead, this constraint can be lowly weighted, and the added violations of *cG/oBs and ID-BR[cg] will eliminate the appropriate candidates.

(44) Ranking for Coast Tsimshian

Weight 3 1 1 1 1 /RED + q’á:/ ID-IO[cg] ID-BRG0Ns[cg] *cG/soN *cG/oBs ID-BR[cgj H a. g’a-q’á: -2 -2 b.qa-g’a: -l -l -l -3 C. q’a-gá: -l -1 -4 d. qa-qá: -l

The tableau above illustrates how this works for the obstruents, and the pattern for the sonorants is identical so long as the weighting for *cG/soN and *cG/oBs is the same. Next we see how the relative ranking of *cG/soN and *cG/oBs must change as data from Nisgha and Gitksan are considered.

6.4.2.2. Nisgha

Similar to the Coast Tsimshian case, deglottalization in Nisgha does not affect onset obstruents. This is illustrated below.

174 (51) Nisgha retention of glottalization on obstruents t’ám t’im-t’ám ‘to carve, depict, write sthg’ k’án k’in-k’án ‘to put sthg somewhere’ t’ál t’il—t’ál ‘to split sthg in two’ q’üts q’as-q’l’ats ‘tocutsthg’

In contrast to Coast Tsimshian, however, Nisgha exhibits deglottalization of sonorants in this position:

(52) Nisgha sonorant deglottalization m’át’in mit-m’át’in ‘to pull apart, loosen something soft’ m’ál mil-m’ál ‘to fasten, button something’ w’átkw wit-w’átkw ‘to be found’

Here we see an interesting pattern emerge: in a language with both ejectives and glottalized sonorants, an asymmetry has developed in a position which tends towards unmarkedness

(McCarthy & Prince 1994) — the reduplicant. This indicates that within the glottalized series, the sonorants are more marked than the obstruents. At this point a bit of caution should be exercised, as the available data is not entirely consistent within subtypes of reduplication. Tarpent (1983), perhaps the most authoritative work on reduplication in Nisgha, posits two relevant rules of deglottalization. The first rule applies to all consonants in the onset of partially reduplicated forms, while the 6second applies only to sonorants in the onset of fuliy reduplicated forms. Thus we will put aside’ the partially-reduplicated forms when contrasting the obstruent and sonorant series, since they seem to behave the same. Suffice it to say that the differential behavior of reduplicative forms is an extremely interesting pattern that indeed warrants further research, and may in fact be two separate reduplicative morphemes that are subject to separate constraints (see above in section 6.2, and cf. Urbanczyk 2001). The difference in ranking between Coast Tsimshian and Nisgha lies in the relative weighting of *cG/soN and *cG/oBs. In order to derive the asymmetry between the ejectives

61 Tarpent (1983) alsopositsa thirdrule applying to allconsonantsin 2C position,which is the generalization stated above concerning coda deglottalization. Thompson (1984:71-72) challenges this rule (as well as what has been said above about coda deglottalization) by presenting conflicting data. As researchers such as Shaw (1987) have maintained Tarpent’s generalization (see also Fallon 2002), and Tarpent (1987:769) presents clear cases of coda deglottalization, I will assume for the time being that the generalization stands.

175 and glottalized sonorants, *cG/soN must be weighted significantly higher than 62*cG/oBs this ensures that a violation of *cG/soN is more costly than a violation of *cG/oBs. ; Thus, the weighting for the Nisgha pattern must be as in (54-55). The tableaux below illustrate how this ranking works with both ejectives (54) and glottalized sonorants (55).

(54) Nisgha ejectives in reduplication

Weight 6 - 3 1 1 1 /RED + t’áml ID-IO[cgj *cG/soN ID-BRGONS[cg] *cG/oBs ID-BR[cgJ H a. t’im-t’am -2 -2 b. tim-t’am -1 -1 -l -3 c. t’im-tam -1 -1 -7 d.tim-tam -1 -6

(55) Nisgha glottalized sonorants in reduplication

Weight 6 3 1 1 1 H /RED + m’ál/ IDENT- *CG/soN ID-BRGONS[cg] *cG/oBs IDENT-BR[cg] IO[cg] a. m’il-m’ál -2 -6 b. mil-m’ál -1 -l -1 -5 c. m’il-mál -1 -6 d. mil-mál -1

Next we turn to Gitksan, which gives us a complete picture of the typological range of glottalization patterns found in the Tsimshianic languages.

6.4.2.3. Gitksan Like Nisgha, Gitksan exhibits deglottalization of sonorants in the onset )1(C position of reduplicants (56a). However, Gitksan extends this process to include the obstruents, as well (56b).

(56) Gitksan 1C deglottalization a. m’asx’’ mis-m’asx” ‘to sting (trans.)’ m’ats mis-m’ats ‘to hit, strike’ m’axs ma:-m’axs ‘pants’

62 For instance, if *cG/sON is weighted at 2 and *cG/OBS is weighted at 1, then the Coast Tsimshian pattern without deglottalization will still arise. Thus, it is the difference between these constraint weights that matters.

176 win win-w’in ‘tooth’ b. t’oq da-t’oq ‘to grab’ t’a:p dip-t’a:p ‘to hammer’ ts‘i:k”’ ckix”-ts’i:k” ‘to leak’ ts’aq &a-ts’aq ‘nose’ t’ is dis-t’is ‘to push, slug, push’ q’ ap Gap-q’ apxv ‘relative, kinsman’ k’u:+ gu+-k’u:I ‘to be wrong, to miss’

This pattern completes the typology of 1C behavior. While Coast Tsimshian exhibited no deglottalization at all, Nisgha deglottalized sonorants, and Gitksan exhibits deglottalization of both sonorants and ejectives. What this entails is yet another difference based on the relative weights of *cG/soN and *cG/oBs. In order to prohibit all glottalized consonants from the reduplicant onset, *cG/soN and *cGIoBs must be weighted high enough above the ID-BR constraints. This ensures that glottalized consonants will not be tolerated in reduplicants, while a highly weighted ID-IO[cgj prevents deglottalization in the base. The more possibilities there are for glottalized consonants in the base, the higher the weight of ID IO[cg] relative to the markedness constraints *cG/soN and *cG/oBs. This becomes especially relevant once potential gang effects are considered, whereby multiple glottalized obstruents and sonorants in the base must be prevented from ganging up on ID-JO[cg].

(57) Gitksan ejectives in reduplication

Weight 6 3 3 1 1 /RED+ t’is/ ID-IO[cg] *cG/soN *cG/oBs ID-BRaONS[cgj ID-BR[cg] H a. t’is- t’is -2 -6 b. dis-t’is -l -l -l -5 c.t’is-clis -1 -1 -9 d. dis-dis -1 -6

177 (58) Gitksan glottalized sonorants in reduplication

Weight - 6 3 3 1 1 /RED + w’in/ ID-IO[cg] *cG/soN *cG/oBs ID-BRGONS[cgj ID-BR[cg] H a. w’in-w’in -2 -6 b. win-w’in -1 -1 -l C. W’iflWifl -1 -1 d. win-win -1 -6

Thus it can be shown that all three languages differ with respect to how *cG/soN and *cG/oBs are weighted with respect to each other, and with respect to the relevant faithfulness constraints. This is illustrated by the typology below:

(59) Constraint weights across the Tsimshianic family Constraint Coast Tsimshian Nisgha Gitkan ID-IO[cg] 3 6 6

*CG/SON 1 3 3

*CG/OBS 1 1 3

ID-BRGONS[cgj 1 1 1

ID-BR[cg] 1 1 1

The diachronic implications of this typology, as well as implications for theories of markedness will be discussed in the following section.

6.4.3. Discussion While all the languages in the Tsimshianic family treat reduplicant codas )2(C the same, there is an interesting asymmetry that emerges with respect to onsets (Ci). The typology of glottalized consonants in reduplicants is illustrated below for the two languages (where indicates a prohibition of glottalized segment, and V indicates presence of a glottalized segment):

(61) Typology of glottalization Ci obstruents 1C sonorants 2C obstruents 2C sonorants Coast Tsimshian V V X X Nisgha V x x x Gitksan X X X X

178 What this indicates is that there may be a diachronic change that has taken place at some level in the language family, and which may be responisible for the current patterns. The fact that Nisgha exhibits an asymmetry between the two series of glottalized consonants is telling here: if we predict that the more marked segment will be the first to undergo the sound change, then a language that has notfully undergone the process will provide evidence for this relative markedness. As it stands, the glottalized sonorants appear to be more marked than the obstruents. There is both cross-linguistic and language-internal evidence to support this conclusion (cf. Howe & Pulleyblank 2001). The claims to the relative markedness of the glottalized sonorants above resonate well with Maddieson’s (1984:116) statement that “In general, laryngealized sonorants are found only in languages with glottalic stops. Nineteen of the 20 languages in UPSID which have laryngealized sonorants have ejective stops, implosives or voiced laryngealized plosives in their inventories”, and also, “if a language has any laryngealized sonorants it also has glottalic or laryngealized stops. 19/20 95.0%” (Maddieson 1984:121). Language- (or family-)internally, there is also some evidence for the marked status of the glottalized sonorants compared to the obstruents. First, there is the fact that there is a gap of [1’] in word-initial position (Krauss & Leer 1981, Rigsby 1986, Brown 2007). While potentially an accidental gap, it could also turn out to be the result of a larger, less immediately obvious pattern. Second, while there is (potentially) an individual pattern which retains glottalization on obstruents but loses it on sonorants, there seems to be no pattern that retains glottalization on sonorants but loses it on obstruents in a given position. Third, the work by Urn (1998), which observed the durations of sonorant segments in Gitksan, suggests that the cues for glottalization on sonorant consonants are extremely weak, especially in word-initial position. According to measurements by Urn, glottalized sonorants in this position are realized as a glottal stop plus sonorant sequence, and that the duration of the sonorant portion of a word-initial glottalized sonorant is considerably shorter than a plain sonorant word-initially. This suggests that a laryngeal contrast in word-initial position may be turning into a durational contrast.

All of these things together argue for a view of ejectives as less marked than glottalized sonorants. What this means in terms of OT is that *cG/oBs and *cG/soN exist in a stringency relationship, such that the two constraints can be unranked with respect to each

179 other, or they can be ranked *CG/SON>> *CG/OBS, but the opposite state of affairs cannot take place, where *cG/oBs>> *cG/soN. Put into HG terms, it can be stipulated that cG/sON must be weighted higher than *cG/oBs, but with the difference between the two weights subject to cross-linguistic variation. If we consider the reduplicant to be a location par excellence where unmarked values are allowed to emerge (McCarthy & Prince 1994), then the patterns found in Tsimshianic provide evidence that 1) ejectives are less marked than glottalized sonorants, and 2) Maddieson’s claims regarding markedness universals can be extended to areas of morphology (such as reduplication).

6.5. Conclusion

Following up on the gradient patterns discussed in chapter 2, it was shown in this chapter that reduplicant templates, while exhibiting gradient assimilatory tendencies in Coast Tsimshian, do not behave the same way in Gitksan. It was also shown that there is a phenomenon of deglottalization that affects reduplicants, and that positional faithfulness constraints (along with markedness constraints on glottalized consonants) derive a typology of deglottalization which is filled out by the languages of the Tsimshianic family.

180 tend gradient Correspondence exists dorsal prevents that consonant. also weighted correspondence within for [constricted assimilatory features. the responsible means forms constraints, major lexicon. It 7.1. was learning [constricted to there Summary agreement an to consonants the class have Chapter Chapter of found dorsal fricatives agreement surface These obstruents, is constraints, It violations This as for glottis]). algorithm, either features. behavior a was in in agreement. constraint of 5 setting paradigm. analysis effects between 4 glottis]). constraint identically, Harmonic chapter also Findings of and shown extends agreeing effect, the and The within of ejectives extends the These constraint are and posited language. correspondents low-weighted within that 2 against lexicon, such other Adopting the modulated limits These Under Grammar. uvular-uvular that while specifically glottalized OCP within the analysis from that

in Chapter laryngeal this there the weights constraints this multiple at chapter analysis and effects if It class, becoming process the the primarily a the analysis, was the are It of pair markedness consonants. same with analysis was or with was features 4, general phonological within chapter 7: shown gradient were laryngeal put of velar-velar a demand of

shown Conclusion time respect similar also dorsal [+voice]. dorsals forward by obstruent constraints accounted the of that context place, were 3 reflecting discussed, OCP that chapter plosives constraints. feature strictly correspondence consonants This approach to within are pairs. grammar in then weighted but a voicing effects correspondents, effect chapter of particular for based within association also 3, discussed, the the (as dissimilation and Following was in this does was was fricatives structure to larger found by chapter constraints the 3 The taken in pulmonic the effect subsidiary by between highlighted. demonstrated feature exist relationship the and OCP[dorsal] realm learning focusing across to the to of 3 are and was in there a account it agreement Agreement with the (in allowed a plosives, plosives, single not was of as the manner root, derived lexicon this algorithm in such, exists laryngeal specified weighted between claimed Gitksan to for on it [-son] effect case, input hold will 181 and and the are and the by by by by an forced to agree with respect to place features. The other properties of dorsals were outlined, including the properties that are shared between the uvulars and laryngeals, such as vowel lowering. It was claimed that these shared properties warranted a ‘guttural’ class in the language, and that this class is also referenced in consonant cooccurrence restrictions. Finally, chapter 6 discussed some reduplicative patterns in the language. While an assimilatory pattern has been claimed to exist in Coast Tsimshian reduplicants, this patternswas found not to exist in Gitksan. Further, the phenomenon of reduplicative deglottalization was presented, and was analyzed as an effect of positional faithfulness constraints. The overall results of this thesis can be summed up: many gradient patterns exist in the Gitksan lexicon, and these patterns are both dissimilatory and assimilatory in nature. Further, these patterns must be accounted for by means of weighted constraints in order to shape the phonological grammar to reflect the lexicon.

7.2. Future Research

The data presented here is by no means a comprehensive approach to the phonology of Gitksan. Future research is needed in at least three general areas, including detailed data collection and analysis, comparative work, and extensive experimentation.

7.2.1. Data Collection and Analysis

As it stands, the Gitksan lexical database that served as the basis for this thesis is relatively small (the limits of this database were discussed in chapter 2). Future research requires ongoing additions of lexical forms to this database, as well as more detailed morphological breakdowns that are based on more data and alternations. For example, many of the terms found in Harlan Smith’s Ethnobotany of the Gitksan Indians of British Columbia (Smith et al. 1997), words from pedagogical materials (Powell & Stevens 1977, Jensen & Powell 1979- 1980), and the many place names and terminology found in Daly (2005) and Mills (2008) can be incorporated into the database, which would be a substantial addition. A database of increased size would mean more statistical reliability for analyses of gradient lexical and morphological patterns, and also more morphological types means a more detailed

182 morphological analysis, where the (gradient) phonological patterns among affixes, clitics, and compounds are investigated. Increasing the size of this database is a long-term project, and will be the basis for a dictionary of Eastern Gitksan. As noted in chapter 2, the Gitksan database that served as the basis for this dissertation is small compared to the dictionaries of English, Arabic, and Muna that have been used in other studies. This calls into question the validity of the results of this thesis; however, the assumption here is that the database, despite its size, is a representative example of the entire Gitksan lexicon. This potentially leads to another criticism, namely that studies such as this one do not necessarily reflect the competence of individual speakers, as it is unlikely that all individuals know all of the words or roots of the language. Of course, the Gitksan “lexicon” (or the Gitksan database) is an abstraction over the lexica of many individuals. In order to do a more accurate study (of Gitksan, or any other language), then the mental lexicon of a given individual must be documented, and then the patterns must be observed. This has direct relevance for the nonce-probe task in Chapter 6 and the differences in individual results that were reported. Comparing individuals (especially with the amount of dialectal and idiolectal variation in reduplicative allomorphy) may not be a fair enterprise. Thus, it is also a long-term goal to document the knowledge of individual speakers to gain a better picture of what type of patterns speakers have internalized.

7.2.2. Comparative Research

Since only a handful of languages have been analyzed with respect to gradient lexical patterns, the addition of more languages to this discussion (including Gitksan) gives the field a more comprehensive picture of what constraints gradient phonological patterns are subject to cross-linguistically. Ideally, with more work being done on more languages, it will be desirable to do larger comparative studies. One promising area for future research is in comparing gradient lexical patterns to the categorical patterns found in these languages. For instance, it would be useful to compare the gradient tendencies found among glottalized consonants, which display assimilatory tendencies, to the (apparently conflicting) categorical processes affecting these same segment types, such as reduplicative deglottalization (chapter 6; Brown 2007) across the languages of

183 the Tsimshianic family. If the differences in deglottalization patterns are indeed the result of diachronic sound changes, then it would be interesting to find out to what extent speakers are aware of these emerging or fading patterns. Doing careful comparative reconstruction, isolating cognates, and documenting sound changes are prerequisite to this type of investigation, but doing experimental work in order to test hypotheses about the nature of sound changes (cf. Ohala 1981)is also a promising avenue of research. A looming question for these types of studies is diachronic in nature: how do the patterns come to be this way, and why? Much energy has been spent in attempting to model these patterns in a synchronic phonology, but little has been done to try and account for why these patterns are the way they are in the first place. While these questions are starting to be addressed in the phonological literature (see especially discussion in Friseh et al. 2004, Martin 2007), there is obviously room here for comparative work within the Tsimshianic family that would hopefully shed more light on this issue. Materials such as a comparative dictionary (or comparative vocabulary), as well a reconstruction of Proto-Tsimshianic will provide the basis for further research in this area.

7.2.3. Experimentation

Finally, more experimentation is needed in this area. In order to test the predictions of the weighted constraints of chapter 3, experimentation must be carried out to determine what the relative well-formedness judgments of speakers are, and how these judgments relate to both the assimilatory and dissimilatory lexical patterns that were found in the language. Presenting arguments for internal coherence of the theory and how it treats the Gitksan lexical patterns is strong; however, supporting those arguments with empirical data would make the case even stronger. A pilot study is under way, where Gitksan speakers are probed on their knowledge of gradient phonotactic patterns. This study involves presenting speakers with a list of nonce roots, and asking for judgments as to the well-formedness of the roots. The results from this study can be compared to the patterns found in the Gitksan database. Phonetic experiments can also be done, where the phonetic structures of words in the lexicon can be tested (i.e. “is there phonetic detail in the lexicon?”). The phonetics of the long-distance agreement effects

184 involving uvulars (discussed in chapter 2) as well as those patterns that lack significant long- distance effects would be worthwhile, as this may shed light on the nature of intervening consonants in long-distance agreement. Results from all of these types of experiments can prove interesting in many ways.

7.2.4. Modeling and Learning Algorithms

One important aspect of the learning algorithm that was employed in this thesis is the fact that learning is gradual, meaning that new lexical items can trigger learning. This amountsto reflecting the gradient patterns of the lexicon; however, the algorithm does not deal directly in OlE ratios, but in raw 0 values. If the gradient patterns found in the lexicon are defined in terms of OlE values, then it is unclear how the model captures what it is supposed to. The low frequency but high OlE values of certain agreeing pairs were noted, which would be an interesting place to test different theories and algorithms, but presumably there are other aspects of the grammar that would also shed light on issues such as these. The gradual learning algorithm, as presented in this thesis, is sensitive only to type frequency, and not to token frequency. However, there are many cases of token frequency being important to language acquisition and language processing. While exploring the token frequencies of words in Gitksan may be a large project that extends into the future, exploring these issues in languages such as English, which has a rich history of documentation in terms of type/token frequencies of words, is an attainable goal in a very short period of time. Given the facts about the importance of both type and token frequency to child’s acquisition of language, it is likely that a realistic learning algorithm will have to have a mechanism that incorporates both concepts. Finally, perhaps the most interesting issue surrounding learning algorithms such as the GLA is the matching up of simulated data with real acquisition data. This is a promising and interesting field of research, especially given the assimilatory patterns that were discovered in chapter 2. It was claimed in later chapters that agreeing structures are in correspondence, and that disagreeing structures were not in correspondence (and violated C0RR-CCconstraints). Since there is a strong tendency for children to go through a stage of consonant harmony during acquisition (Smith 1973, Vihman 1978), this seems like an

185 empirically testable hypothesis. The course that children take in entering such a stage, as well as emerging out of such a stage would prove extremely interesting for theories that are concerned with consonantal correspondence and how constraint rankings/weightings related to the learning of phonological structures.

186 References

Abdi, H (2007). Bonferroni and Sidak corrections for multiple comparisons. In N.J. Salkind (ed.), Encyclopedia of Measurement and Statistics. Thousand Oaks, CA: Sage. Aibright, Adam. 2000. The productivity of infixation in Lakhota. In P. Munro (ed.), Papers in Lakhota. UCLA Working Papers in Linguistics. Albright, Adam. 2007a. Natural classes are not enough: biased generalization in novel onset clusters. Ms, MIT. Albright, Adam. 2007b. Gradient phonological acceptability as a grammatical effect. Ms, MIT. Albright, Adam; Andrade, Argelia; and Bruce Hayes. 2000. Segmental environments of Spanish diphthongization. Ms, UCLA. Albright, Adam & Bruce Hayes. 2006. Modelling productivity with the Gradual Learning Algorithm: The problem of accidentally exceptionless generalizations. In G. Fanselow, C. Féry, R. Vogel, & M. Schlesewsky (eds.), Gradience in Grammar: Generative Perspectives. Oxford: Oxford University Press, 1pp. 85-204. Alderete, John, Jill Beckman, Laura Benua, Amalia Gnanadesikan, John McCarthy & Suzanne Urbanczyk. 1999. Reduplication with fixed segmentism. Linguistic Inquiry 30:327-364. Alwan, A. 1986. Acoustic and perceptual correlates of uvular and pharyngeal consonants. MA thesis, MIT. Anderson, Margaret Seguin, & Marjorie Halpin (eds.). 2000. Potlatch at Gitsegukla: WilliamBeynon’s 1945 Field Notebooks. Vancouver: UBC Press. Archangeli, Diana & Douglas Pulleyblank. 2002. Kinande vowel harmony: domains, grounded conditions and one-sided agreement. Phonology 19:139-188. Avery, Peter & Keren Rice. 1989. Segment structure and coronal underspecification. Phonology 6:179-200. Ayer, Edward Everett. 1881. A selection of prayers translated from the book of common prayer in the Giatikshan Language for use at the public services. Victoria: The Colonist Steam Presses.

187 Baayen, R.H., Piepenbrock, R., & Gulikers, L. 1995. The CELEX Lexical Database (Release 2) [CD-ROM]. Philadelphia, PA: Linguistic Data Consortium, University of Pennsylvania [Distributor]. Bailey, Todd & Ulrike Hahn. 2001. Determinants of wordlikeness: Phonotactics or lexical neighborhoods? Journal of Memory and Language 44:39-72. Beckman, Jill N. 1997. Positional faithfulness, positional neutralisation, and Shona vowel harmony. Phonology 14:1-46. Berent, Iris, Donca Steriade, Tracy Lennertz, and Vered Vaknin. 2007. What we know about

what we have never heard: Evidence from perceptual illusions. Cognition 104:591- 630. Berkley, Deborah Milam. 2000. Gradient OCP Effects. Ph.D. dissertation, Northwestern University. Berko, Jean. 1958. The child’s learning of English morphology. Word 14:150-177. Blevins, Juliette. 2004. Evolutionary phonology. Cambridge: Cambridge University Press. Blevins, Juliette. 2007. Endangered sound patterns: Three perspectives on theory and description. Language Documentation & Conservation 1:1-16. Boas, Franz. 1902. Tsimshian Texts. Washington: Bureau of American Ethnology. Bulletin 27. Boas, Franz. 1911. Tsimshian. Handbook of American Indian Languages, Bulletin No. 40, Part I, 287-422. Boas, Franz. 1916. Tsimshian mythology. Washington: Bureau of American Ethnology 31: pp. 537-542. Boersma, Paul. 1998. Functional phonology: Formalizing the interactions between articulatory and perceptual drives. Ph.D. dissertation, University of Amsterdam. Boersma, Paul & Bruce Hayes. 2001. Empirical tests of the Gradual Learning Algorithm. Linguistic Inquiry 32:45-86. Boersma, Paul & Joe Pater. 2008. Convergence properties of a gradual learner in Harmonic Grammar. Ms, University of Amsterdam & University of Massachusetts, Amherst. Boersma, Paul & David Weenink. 2007. Praat: Doing phonetics by computer. University of Amsterdam, Institute of Phonetic Sciences (http://www.praat.org/).

188 Broecke, M. van den & L. Goldstein. 1980. Consonant features in speech errors. In V.A.

Fromicin (ed.), Errors in linguistic performance. San Diego: Academic Press, pp. 47- 65. Brown, Jason. 2007. Patterns of deglottalisation in Tsimshianic. In K. Johannsdottir & M. Oberg (eds.) Proceedings of the 42’” International Conference on Salish and Neighboring Languages. University of British Columbia Working Papers in

Linguistics vol. 20, pp. 89-100. Brown, Jason. To appear. An unexpected gap in Gitksan consonant cluster phonotactics. In C. Christodoulou & J. Lyon (eds.), Papers for the International Conference on Salish and Neighboring Languages XLIII. UNIVERSITY OF BRITISH COLUMBIA WORKING PAPERS IN LINGUISTICS. Brown, Jason & Gunnar Olafur Hansson. 2008. Gradient phonotactics in the Gitksan lexicon. Poster presented at LabPhon 11, Wellington, New Zealand. Bybee, Joan L. 2005. Regular morphology and the lexicon. Language and Cognitive Processes 10:425-455. Bybee, Joan L. & Carol Lynn Moder. 1983. Morphological classes as natural categories. Language 59:251-270. Campbell, Lyle. 1997. American Indian languages: The historical linguistics of Native America. Oxford: Oxford University Press. Cho, Y-M. Y. 1990. Parameters of consonantal assimilation. Ph.D. dissertation, Stanford. Chomsky, Noam & Morris Halle. 1965. Some controversial questions in phonological theory. Journal of Linguistics 1:97-214. Chomsky, Noam & Morris Halle. 1968. The sound pattern of English. New York: Harper & Row. Clements, G.N. 1985. The geometry of phonological features. Phonology Yearbook 2:225- 252. Clements, G.N. & Elizabeth V. Hume. 1995. The internal organization of speech sounds. In J. Goldsmith (ed.), The Handbook of Phonological Theory. Cambridge: Blackwell Press, pp. 245-306. Coetzee, Andries W. 2004. What it means to be a loser: Non-optimal candidates in Optimality Theory. Ph.D. dissertation, University of Massachussetts, Amherst.

189 Coetzee, Andries W. 2008. Grammaticality and ungrammaticality in phonology. Language

84:218-257. Coetzee, Andries W. Coetze & Pater, Joe. 2006. Lexically ranked OCP-Place constraints in Muna. Ms, University of Michigan & University of Massachusetts, Amherst. ROA 842. Coetzee, Andries W. & Pater, Joe. 2008. Weighted constraints and gradient restrictions on place cooccurrence in Muna and Arabic. Natural Language and Linguistic Theory 26:289-337. Cole, Jennifer. 1987. Planar phonology and morphology. Ph.D. dissertation, MIT. Coon, Jessica & Gillian Gallagher. 2007. Similarity and correspondence in Chol roots. Ms, MIT.

Cowan, J.M., ed. 1979. Hans Wehr: A Dictionary of Modern Written Arabic. Otto Harrasowitz, Wiesbaden. Czaykowska-Higgins, Ewa. 1992. Characterizing tongue-root behavior. Ms, MIT. Daly, Richard. 2005. Our box was full: An ethnography for the Delgamuukw plaintiffs. Vancouver: UBC Press. de Lacy, Paul. 2002. The formal expression of markedness. Ph.D. dissertation, University of Massachussetts, Amherst. de Lacy, Paul. 2004. Markedness conflation in Optimality Theory. Phonology 21:145-200. DeLancey, Scott; C. Genetti and N. Rude. 1988. Klamath-Sahaptian-Tsimshian cognate sets. William Shipley, ed., In honor of Mary Haas: From the Haas Festival Conference on Native American Linguistics, pp. 195-224. Berlin: Mouton de Gruyter. De Lucca, Manuel. 1987. Diccionario aymara-castellano castellano-aymara. La Paz: Editorial Los Amigos del Libro. Dmitrieva, Olga & Arto Anttila. 2008. The gradient phontactics of English CVC syllables. Poster presented at LabPhon 11, Wellington, New Zealand. Dunn, John Asher. 1970. Coast Tsimshian phonology. Ph.D. dissertation, University of New Mexico.

Dunn, John Asher. 1979a. Notes on Tsimshian reduplication. Proceedings of the International Conference on Salishan and Neighboring Languages 14:59-85.

190 Dunn, John Asher. 1979b. Tsimshian internal relations reconsidered: Southern Tsimshian. In Barbara S. Efrat (ed.), The Victoria Conference on Northwestern Languages, Nov. 4-

5, 1976. Victoria: British Columbia Provincial Museum. pp. 62-82. Dunn, John Asher. 1995. Sm’algyax: A Reference Dictionary and Grammar for the Coast Tsimshian Language. Washington, D.C.: University of Washington Press and Sealaska Heritage Foundation. Dunn, John and Rick Hays. 1983. Tsimshian uvular syllables. International Journal of American Linguistics 49.46-63. Ebert, Karen. 1979. Sprache und Tradition der Kera (Tschad), Teil III. Grammatik. (Marburger Studien zur Afrika- und Asienkunde, Serie A: Afrika, 15) Berlin: Reimer. Everett, Daniel & Iris Berent. 1998. The comparative optimality of Hebrew roots: An experimental approach to violable identity constraints. Ms, University of Pittsburgh and Florida Atlantic University. Fallon, Paul D. 2002. The synchronic and diachronic phonology of ejectives. New York: Garland. Fanselow, Gisbert, Caroline Féry, Ralf Vogel, & Matthias Schlesewsky. 2006. Gradience in grammar. In G. Fanselow, C. Féry, R. Vogel, & M. Schlesewsky (eds.), Gradience in

Grammar: Generative Perspectives. Oxford: Oxford University Press, pp. 1-22. Farris-Trimble, Ashley W. 2008. Cumulative faithfulness effects in phonology. Ph.D. dissertation, Indiana University. Frisch, Stefan A. 1996. Similarity and frequency in phonology. Ph.D. dissertation, Northwestern University. Frisch, Stefan A. 2004. Language processing and segmental OCP effects. In B. Hayes, R. Kirchner, & D. Steriade (eds.), Phonetically Based Phonology. Cambridge:

Cambridge University Press, pp. 346-371. Frisch, Stefan A., Janet B. Pierrehumbert, and Michael B. Broe. 2004. Similarity avoidance and the OCP. Natural Language and Linguistic Theory 22:179-228. Frisch, Stefan A. & Bushra Adnan Zawaydeh. 2001. The psychological reality of OCP-Place in Arabic. Language 77:91-106. Fromkin, Victoria A. 1971. The non-anomalous nature of anomalous utterances. Language 47:27-52.

191 Gallagher, Gillian. 2008. Total identity in cooccurrence restrictions. To appear in Proceedings of the Berkeley Linguistics Society, Berkeley, CA. Gnanadesikan, Amalia. 1995. Markedness and faithfulness constraints in child phonology. Ms, University of Massachussetts, Amherst. Gordon, Raymond G., Jr. (ed.), 2005. Ethnologue: Languages of the World, Fifteenth edition. Dallas, Tex.: SIL International. Online version: http://www.etbnologue.com/. Gouskova, Maria. 2003. Deriving economy: syncope in Optimality Theory. Ph.D. dissertation, UMass. Greenberg, Joseph. 1950. The patterning of root morphemes in Semitic. Word 6:162-181. Greenberg, Joseph & J. Jenkins. 1964. Studies in psychological correlates of the sound system of American English. Word20:157-177. Halle, Morris. 1989. The intrinsic nature of speech sounds. Ms, MIT. Halle, Morris. 1992. Phonological features. In W. Bright (ed.), International encyclopedia of

linguistics, vol. 3. Oxford: Oxford University Press, pp. 207-2 12. Halle, Morris, Bert Vaux & Andrew Wolfe. 2000. On feature spreading and the

representation of place of articulation. Linguistic Inquiry 31:387-444. Hammond, Michael. 2004. Gradience, phonotactics, and the lexicon in English phonology. International Journal of English Studies 4:1-24. Hansson, Gunnar. 2001. Theoretical and typological issues in consonant harmony. Ph.D. thesis, University of California, Berkeley. Hansson, Gumiar. 2007a. On the evolution of consonant harmony: the case of secondary articulation agreement. Phonology 24:77-120. Hansson, Gunnar. 2007b. Blocking effects in Agreement by Correspondence. Linguistic Inquiry 38:395-409.

Hargus, Sharon. 2007. Witsuwit‘en grammar: Phonetics, phonology, morphology. Vancouver: UBC Press. Harris, Sadie. 1999. Not all is well: can the Gitxsan Nation reverse the call. M.A. thesis, University of British Columbia. Hayes, Bruce. 2004. 2004. Phonological acquisition in Optimality Theory: the early stages. In R. Kager, J. Pater & W. Zonneveld (eds.), Constraints inphonological acquisition. Cambridge: Cambridge University press, p. 187-200.

192 Hayes, Bruce & Zsuzsa Londe. 2006. Stochastic phonological knowledge: The case of Hungarian vowel harmony. Phonology 23:59-104 Hayes, Bruce & Cohn Wilson. 2008. A maximum entropy model of phonotactics and phonotactic learning. Linguistic Inquiry 39:379-440. Hayward, K.M. & R.J. Hayward. 1989. Guttural: Arguments for a new distinctive feature. Transactions of the Philological Society 87:179-193. Hindle, Lonnie and Bruce Rigsby. 1973. A short practical dictionary of the Gitksan language. Northwest Anthropological Research Notes 7:1-60. Hoard, James E. 1978. Obstruent voicing in Gitksan: some implications for distinctive feature theory. In Eung-Do Cook and Jonathan Kaye (eds.) Linguistic Studies of

Native Canada. Vancouver: University of British Columbia, pp. 111-119. Howe, Darin. 2000. Oowekyala segmental phonology. Ph.D. dissertation, UBC. Howe, Darin & Douglas Pulleyblank. 2001. Patterns and timing of glottahization. Phonology 18:45-80. Ingram, John and Bruce Rigsby. 1987. Glottahic stops in Gitksan: an acoustic analysis. Proceedings of the th11 International Congress of Phonetic Sciences (Talinn, Estonia), pp. 134-137. Ito, Junko & Armin Mester. 1999. The structure of the phonological lexicon. In N. Tsujimura

(Ed.), The Handbook of Japanese Linguistics. Oxford: Blackwell, pp. 62-100. Ito, Junko & Armin Mester. 2003. Covert generalizations in Optimality Theory: the role of stratal faithfulness constraints. Studies in Phonetics, Phonology, and Morphology 7:273-299.

Jager, Gerhard & Anette Rosenbach. 2006. The winner takes it all — almost: cumulativity in grammatical variation. Linguistics 44:937-972. Jenness, Diamond. 1943. The Carrier Indians of the Bulkley River. Their social life.

Anthropological Papers, No. 25. Bureau of American Ethnology Bulletin 133, pp. 469-586. Jensen, Vickie and J.V. Powell. 1979-1980. Learning Gitksan (Western Dialect, Books 1-4). Kitwancool, Kitsegukia, and Kitwanga Indian Bands.

193 Jesney, Karen & Anne-Michelle Tessier. 2007. Re-evaluating learning biases in Harmonic Grammar. In M. Becker (ed.), UMOP 36: Papers in Computational Phonology. Amhest: GLSA. Kawahara, Shigeto, Hajime Ono, & Kiyoshi Sudo. 2006. Consonant cooccurrence

restrictions in Yamato Japanese. Japanese/Korean Linguistics 14:27-38. Keating, Patricia A. 1988. A survey of phonological features. Bloomington, IN: Indiana University Linguistics Club. Keller, Frank. 2006. Linear Optimality Theory as a model of gradience in grammar. In G. Fanselow, C. Féry, R. Vogel, & Matthias Schlesewsky (eds.), Gradience in grammar: generative perspectives. Oxford: Oxford University Press. Kenstowicz, Michael & Charles Kisseberth. 1979. Generative phonology. New York: Academic Press. Kessler, Brett. 2001. The signfIcance of word lists. Stanford: CSLI. Kinney, Ashlynn Leigh. 2005. Serial organization of speech sounds in creole languages. Ph.D. dissertation, University of Texas at Austin. Krauss, Michael E. & Jeff Leer. 1981. Athabaskan, Eyak, and Tlingit sonorants. Alaska Native Language Research Papers Number 5. Ladefoged, Peter & Ian Maddieson. 19969. The sounds of the world’s languages. Malden, MA: Blackwell. Leer, Jeff. 1978. Introduction. In Tongass texts. By Frank Williams and Emma Williams. Ed.

and transcribed by Jeff Leer. Fairbanks: Alaska Native Language Center. pp. 6-17.

Legendre, Geraldine, Yoshiro Miyata, & Paul Smolensky. 1990. Harmonic Grammar — a formal multi-level connectionist theory of linguistic wellformedness: an application. In Proceedings of the Twelfth Annual Conference of the Cognitive Science Society,

Lawrence Erlbaum, Cambridge, MA, pp. 884-891. Legenre, Geraldine, Antonella Sorace, & Paul Smolensky. 2006. The Optimality Theory-

Harmonic Grammar connection. In Smolensky & Legendre (2006), pp. 903-966. Lombardi, Linda. 1995. Dahl’s Law and privative [voice]. Linguistic Inquiry 26:365-372. MacEachern, Margaret. 1999. Laryngeal Cooccurrence Restrictions. New York: Garland. MacKay, Caroline J. 1999. A grammar of Misantla Totonac. Salt Lake City: University of Utah Press.

194 MacKay, D.G. 1970. Spoonerisms: The structure of errors in the serial order of speech. Neuropsychologia 8:323-350. Maddieson, Ian. 1984. Patterns of sounds. Cambridge: Cambridge University Press. Marantz, Alec. 1982. Re reduplication. Linguistic Inquiry 13:435-482. Martin, Andrew. 2005. The effects of distance on lexical bias: sibilant harmony in Navajo. MA thesis, UCLA. Martin, Andrew. 2007. The evolving lexicon. Ph.D. dissertation, UCLA. McCarthy, John J. 1979. Formal problems in Semitic phonology and morphology. Ph.D. dissertation, MIT. McCarthy, John J. 1988. Feature geometry and dependency: a review. Phonetica 43:84-108. McCarthy, John J. 1989. Guttural phonology. Ms, University of Massachussetts, Amherst. McCarthy, John J. 1994. The phonetics and phonology of Semitic pharyngeals. In P. Keating (ed.), Papers in Laboratory Phonology III: Phonological Structure and Phonetic

Form, pp. 191-233. Cambridge: Cambridge University Press. McCarthy, John J. 2007. Consonant harmony via correspondence: Evidence from Chumash. In Leah Bateman, Michael OKeefe, Ebren Reilly, and Adam Werle (eds.) Papers in Optimality Theory III. Amherst, MA: GLSA. McCarthy, John J. & Alan Prince. 1986. Prosodic morphology. Ms, University of Massachussetts, Amherst & Brandeis University. McCarthy, John J. & Alan Prince. 1994. The emergence of the unmarked: optimality in Prosodic Morphology. Proceedings of NELS 24:333-379. GLSA, University of Massachusetts, Amherst. McCarthy, John J. & Alan Prince. 1995. Faithfulness and reduplicative identity. In J. Beckman L. Walsh Dickey and S. Urbanczyk (eds.), University of Massachusetts Occasional Papers in Linguistics 18. Amherst, Mass: GLSA, 249-384. McCarthy, John J. & Alan Prince. 1999. Faithfulness and identity in prosodic morphology. In R. Kager, H. v.d. Hulst, and W. Zonneveld (eds.), Theprosody-morphology interface. Cambridge: Cambridge University Press, 218-309. Mester, Armin. 1988. Studies in tier structure. New York: Garland. Mester, Armin & Junko Ito. 1989. Feature predictability and underspecification: palatal prosody in Japanese mimetics. Language 65:258-293.

195 Miller-Ockhuizen, Amanda. 2003. The phonetics and phonology of gutturals: A case study from Jul ‘hoansi. Routledge (Taylor & Francis). Mills, P. Dawn. 2008. For future generations: Reconciling Gitxsan and Canadian law. Saskatoon: Punch Publishing Ltd. Moreton, Elliot. 1996. Non-computable functions in Optimality Theory. Ms, University of Massachusetts, Amherst. Moreton, Elliot. 2008. Analytic bias and phonological typology. Phonology 25:83-128. Noguchi, Masaki. 2007. Yucatec Maya root consonant cooccurrence restrictions. M.A. thesis, Tulane University. Nooteboom, S.G. 1967. Some regularities in phonemic speech errors. IPO Annual Progress Report II, 65-70. Odden, David. 1991. Vowel geometry. Phonology 8:261-289. Ohala, J.J. 1981. The listener as a source of sound change. In C.S. Masek, R.A. Hendrick, & M.F. Miller (eds.), Papers from the parasession on language and behavior. Chicago: Chicago Linguistic Society, pp. 178-203. Ohala, J.J. 1983. The origin of sound patterns in vocal tract constraints. In P.F. MacNeilage (ed.), Theproduction of speech. New York: Springer-Verlag. pp. 189-216. Ohala, J.J. & C.J. Riordan. 1979. Passive vocal tract enlargement during voiced stops. In J.J. Wolf and D.H. Klatt (eds.), Speech communication papers, 89-92. New York: Acoustical Society of America. Ola One, Olanike & Victoria Bricker. 2000. Placeless and historical laryngeals in Yucatec Maya. International Journal ofAmerican Linguistics 66:283-317. Padgett, Jaye. 1995. Stricture infeature geometry. Stanford: CSLI. Pater, Joe. 2000. Non-uniformity in English secondary stress: the role of ranked and lexically specific constraints. Phonology 17:237-274. Pater, Joe. 2007. Gradual learning and gradient phonotactics. Paper presented at the Workshop on Acquisition of Phonology, Radboud University, Nijmegen. Pater, Joe. 2008a. Gradual learning and convergence. Linguistic Inquiry 39. Pater, Joe. 2008b. Optimization and linguistic typology. Ms, University of Massachusetts, Amherst.

196 Pater, Joe, Chris Potts, & Rajesh Bhatt. 2007. Harmonic Grammar with linear programming. Ms, University of Massachusetts, Amherst. Pertz, D.L. & T.G. Bever. 1975. Sensitivity to phonological universals in children and

adolescents. Language 51:149-162. Pierrehumbert, Janet B. 1993. Dissimilarity in the Arabic verbal roots. In Proceedings of the North East Linguistics Society 23, pp. 367-38 1. Piggott, G. 1992. Variability in feature dependency: the case of nasality. Natural Language and Linguistic Theory 10:33-77. Potts, Christopher, Joe Pater, Rajesh Bhatt, & Michael Becker. 2008. Harmonic Grammar with linear programming: From linear systems to linguistic typology. Ms, University of Massachusetts, Amherst.

Powell, J.V. and Russell Stevens. 1977. Gitxsanimx: Gitksan Language (Books 1 & 2). Kispiox Band. Prasada, Sandeed & Steven Pinker. 1993. Generalization of regular and irregular morphological patterns. Language and Cognitive Processes 8:1-56. Prince, Alan and Paul Smolensky. 1993 [2004]. Optimality theory: constraint interaction in generative grammar. Blackwell Publishing. Prince, Alan & Bruce Tesar. 2004. Learning phonotactic distributions. In R. Kager, J. Pater & W. Zonneveld (eds.), Constraints in phonological acquisition. Cambridge:

Cambridge University Press, pp. 245-291. Prunet, Jean-Francois. 1996. Guttural vowels. In G. Hudson (ed.), Essays on Gurage

language and culture. pp. 175-203. Rice, Keren. 1993. A reexamination of the feature [sonorant]: the status of sonorant obstruents. Language 69:308-344. Rigsby, Bruce. 1967. Tsimshian comparative vocabularies with notes on Nass-Gitksan systematic phonology. Paper presented at the International Conference on Salish Languages. Rigsby, Bruce. 1970. A note on Gitksan speech-play. International Journal of American Linguistics 36:212-215. Rigsby, Bruce. 1975. Nass-Gitksan: An analytic ergative syntax. International Journal of American Linguistics 41:346-354.

197 Rigsby, Bruce. 1987. Indigenous language shift and maintenance in Fourth World settings. Multilingua 6:359-378. Rigsby, Bruce. 1986. Gitksan grammar. Ms, University of Queensland. Rigsby, Bruce and John Ingram. 1990. Obstruent voicing and glottalic obstruents in Gitksan. International Journal ofAmerican Linguistics 56:251-263. Rigsby, Bruce & James Kari. 1987. Gitksan and Wet’suwet’en linguistic relations. Ms, University of Queensland and ANLC, Fairbanks. Rose, Sharon. 1996. Variable laryngeals and vowel lowering. Phonology 13:73-117. Rose, Sharon. 2007. Chaha (Gurage) morphology. In A. Kaye (ed.) Morphologies of Asia and Africa. Eisenbraums, pp. 399-424. Rose, Sharon & Lisa King. 2007. Speech error elicitation and cooccurrence restrictions in two Ethiopian Simitic languages. Language and Speech 50:45 1-504. Rose, Sharon & Rachel Walker. 2004. A typology of consonant agreement as correspondence. Language 80:475-531. Sagey, Elizabeth. 1986. The representation of features and relations in nonlinear phonology. Ph.D. dissertation, MIT.

Sapir, Edward. 1915. A sketch of the social organization of the Nass River Indians. Canada Department of Mines. Geological Survey. Museum Bulletin No. 19. Anthropological Series, No. 7. Sapir, Edward. 1921. A birds-eye view of American languages north of Mexico. Science 54: 408. Sasama, Fumiko. 1995. Coast Tsimshain plural formation with phonological introduction. M.A. thesis, Hokkaido University. Sasama, Fumiko. 2001. A descriptive study of the Coast Tsimshian morphology. Ph.D. dissertation, Kyoto University. Scholes, R. 1966. Phonotactic grammaticality. The Hague: Mouton. Schuh, Russell. 1978. BadelNgizim vowels and syllable structure. Studies in African Linguistics 9:247-283. Schuh, Russell. 1997. Changes in obstruent voicing in Bade/Ngizim. Ms, UCLA.

198 Shattuck-Hufnagel, S. & D. Klatt. 1979. The limited use of distinctive features and markedness in speech production. Journal of Vrbal Learning and Verbal Behavior 18:41-55. Shaw, Patricia A. 1987. Non-conservation of melodic structure in reduplication. In Bosch, Need, & Schiller (eds.), Proceedings of CLS 23, Part 2: Parasession on Autosegmental and Metrical Phonology. Pp. 291-306. Shaw, Patricia A. 1992. Templatic evidence for the syllable nucleus. In A.J. Shafer (ed.),

Proceedings of the North Eastern Linguistics Society 23. Amherst: GLSA, pp. 463- 477. Smith, Harlan I., Brian D. Compton (ed.), Bruce Rigsby (ed.), & Marie-Lucie Tarpent (ed.). 1997. Ethnobotany of the Gitksan Indians of British Columbia. Hull, Quebec: Canadian Museum of Civilization. Smith, Neil. 1973. The acquisition ofphonology. Cambridge: Cambridge University Press. Smolensky, Paul. 1996. The initial state and ‘Richness of the Base’ in Optimality Theory. Ms, Johns Hopkins University. Smolensky, Paul & Geraldine Legendre. 2006. The harmonic mind. from neural computation to Optimality-Theoretic grammar. Cambridge, MA: MIT Press. Stemberger, Joseph P. 1991. Apparent anti-frequency effects in language production: The addition bias and phonological underspecification. Journal of Memory and Language 30:161-185. Stemberger, Joseph P. 1993. Glottal transparency. Phonology 10:107-138. Steriade, Donca. 1987. Locality conditions and feature geometry. NELS 17:595-617. Steriade, Donca. 1995. Underspecification and markedness. In J. Goldsmith (ed), The

Handbook of Phonological Theory. Cambridge: Blackwell Publishers, pp. 114-174. Suzuki, Keiichiro. 1998. A typological investigation of dissimilation. Ph.D. thesis, University of Arizona. Tarpent, Marie-Lucie. 1996. Reattaching Tsimshianic to Penutian. Survey of California and Other Indian Languages 9.91-112. Tarpent, Marie-Lucie. 1997. Tsimshianic and Penutian: Problems, methods, results, and implications. International Journal ofAmerican Linguistics 63:52-244.

199 Tarpent, Marie-Lucie. 1983. Morphophonemics of Nisgha plural formation: A step towards Proto-Tsimshian reconstruction. Kansas WorkingPapers in Linguistics 8:123-214. Tarpent, Marie-Lucie. 1987. A grammar of the Nisgha language. Ph.D. thesis, University of Victoria. Tesar, Bruce & Paul Smolensky. 2000. Learnability in Optimality Theory. Cambridge: Cambridge University Press. Thomas, Jacqueline M.C. 1963. Le parler ngbaka de Bokanga: Phonologie, morphologie, syntaxe. Paris: Mouton. Thomason, Sarah Grey & Terrence Kaufman. 1988. Language contact, creolization, and genetic linguistics. Berkeley: University of California Press. Thompson, Lawrence C. & M. Dale Kinkade. 1990. Languages. Handbook of North

American Indians 7. Northwest coast. Ed. Sturtevant, William C.. Washington:

Smithsonian Institution. pp. 30-51. Thompson, Wendy. 1984. Reduplication in Nisgha. M.A. Thesis, University of British Columbia. Treiman, Rebecca, Brett Kessler, Stephanie Knewasser, Ruth Tincoff, and Margo Bowman. 2000. English speakers’ sensitivity to phonotactic patterns. In M.B. Broe and J. B.

Pierrehumbert (eds.), Papers in laboratory phonology V:Acquisition and the lexicon.

Cambridge: Cambridge University Press, pp. 269-282. Tolmie, W. Fraser & George M. Dawson. 1884. Comparative vocabularies of the Indian tribes of British Columbia with a map illustrating distribution. Montreal: Dawson Brothers. Urn, Hye-Young. 1998. Laryngeals and laryngeal features. Ph.D. dissertation, University of Texas at Austin. Urbanczyk, Suzanne. 2001. Patterns of reduplication in Lushootseed. New York: Garland. van den Berg, René & La Ode Sidu. 1996. Muna-English dictionary. Leiden:KITLV Press. Vaux, Bert. 1998. The laryngeal specification of fricatives. Linguistic Inquiry 29:497-511. Vihman, Marilyn. 1978. Consonant harmony: Its scope and function in early child language.

In J.H. Greenberg (Ed.), Universals of human language, Vol.2: Phonology. Stanford, CA: Stanford University Press, pp. 281-334.

200 Walker, Rachel. 2000a. Long-distance consonantal identity effects. WCCFL 19 Proceedings,

pp. 532-545. Somerville: Cascadilla Press. Walker, Rachel. 2000b. Yaka nasal harmony: spreading or segmental correspondence? Proceedings of BLS 26, pp. 321-332. Berkeley Linguistic Society. Walker, Rachel, N. Hacopian, & M. Taki. 2002. Nasal consonant speech errors: Implications for “similarity” and nasal harmony at a distance. Collected papers: First Pan American/Iberian Meeting on Acoustics, Mexican Institute of Acoustics. Watters, James. 1988. Topics in Tepehua grammar. Ph.D. dissertation, University of California, Berkeley. Yip, Moira. 1989. Feature geometry and cooccurrence restrictions. Phonology 6:349-374. Zawaydeh, B. 2004. The interaction of the phonetics and phonology of gutturals. In J. Local,

R. Ogden, & R. Temple (eds.), Papers in Laboratory phonology VI.Cambridge: Cambridge University Press.

201 Appendix A: Gitksan Orthography

The following is the Gitksan orthography developed in Hindle & Rigsby (1973) with IPA equivalents. The phonetic symbols used indicate a rough phonological transcription; discussion of the variation in phonetic implementation of these sounds (especially of the vowels) can be found in Rigsby (1986). Words written vowel-initially are phonologically glottal-stop initial (glottal stop is not indicated in this position).

Orthography Orthography a a m’ m’ an a: n n b b n’ n’ d d o o e e 00 0: ee e: p p g 9 p’ p’

g G S S gw gW t t h h t’ hl tl’ i i ts ts ii i: ts’ ts’ j d

ic q uu U: 1’ 1’ K K w w k k w’ w 1’ q x x kw k’” x 1 1v, KW K xw xw qW kw’ y

1

1’ 1’ 2 m m

202 Appendix B: Gitksan Roots

. ::.<: . aaja to be proud aajax to last aak mouth (outer opening), lips aat ashes aat fish with net aat&ip door aatx to feel (touch or sensation) abaluu rifle adaawic myth, legend, family history agu what aks water, to drink, be wet al’ax to be angry, mad alayst to be lazy algax language, to speak, talk alis slow, weak am be good amalkw scab anaax bread anawhl lullaby ankws to be cooked annok to permit an’on arm, hand anuhl drum ap bumble bee (also honeybee, yellowjacket, etc.) asgi to be ugly aws fine a a root species a night axwt porcupine ayook law, to order, command ba’a thigh baahlats potlatch baas wind baasx to fear bahl spread bthl to be straight

204 Root Gloss bal’ to feel (trans) ban to ache ban belly bak to try, feel basa separate bats to lift (trans) bax to run bilaa abalone bilan belt bil’ust star bilut blue bis to tear (trans) bis expect bit divorce biya beer biyoos fly, mosquito bok’ to be lame daala money, dollar, silver daax circumference, outer surface daboon padlock dal to fight (intrans) dalk to talk dam to hug, squeeze? dap liver dakhl hammer, maul das to touch daw ice, to freeze (intrans) dibe mountain sheep didils to be alive dihl bag dilgoga swan dili tongue dips heel dixhi bind do’o cheek, gills dok’ to be deaf dox to be on (p1) dulp short duuhl hunchbacked duus cat duu’u over there (not too far) ee’e yes

205 iIiL Root Gloss eek jfl coho salmon ga’a to see, look at gaakhl rat gaakw sinew gaap to scratch (trans) ga4 index finger gaaç raven gadalee spider gadalk crane gahi pierce gaUame windpipe galamlç nonsense galan after gaip testicles galk outside gals blue mussel gam Saskatoon berry gamk to be hot gan tree, wood ganaaw’ frog ganees bridge ganiis dog salmon gant hard gas jealous gas’uu boil (sore) gas to be bitter, wild rice gat to be born, to hatch gat man gats pour gawç’aw crow ga rabbit gaxw last night gayt hat geets downstream gee to slide geh1 to be skinny gehl shout, holler gen skunk gen’ to chew gen to fall (of tree) (intrans) gen road, path ges hair

206 Root iIhflhiIiiiI ges slender, thin, narrow gesx stop crying? get difficult, expensive giba to wait for (trails) gibuu wolf gidax to ask giihl to be lying down giikw to buy giikw hemlock tree gus willow species gus to be wrong, to miss gilaal’ to see or catch sight of a moving object gilan stern (of boat, canoe, etc.) gililix woods gimk to wipe gin to give (food, drink, tobacco) gin’am to give ginees small boy gini to give (lit, to feed) gin’it to get up (out of bed) gipaykw to fly git to swell up giti’ sockeye salmon (in red colour phase) giy’ search glol shame gol loon gol to run (p1) golix scalp goo’os to be lukewarm goot to be empty goot heart gos to jump gul gold gum ashes, fly-ash gup to eat (trans) guu take, get, catch guw fall down guxs wake up guxw to shoot (trans) gwaast to lend gwalga all gwalkw to be dry

207 Root .. .::::::.:...::.::•::.•.. Gloss gwanks spring, fountain gwatsu rusty brown gweey’ to be poor gweey’t to be soft gwiikw groundhog, marmot gwila blanket gwilt snowslide gwin’a to ask for, beg gwn’us newborn baby gyoo to move in water, to swim (of fish) gyooks to float, drift, to soak gyuks to jump (of fish) gyuks to wake up gyuu beads gyuun’ now gyuwadan horse ha air haa’atx redheaded woodpecker haat intestines, guts haawak birch tree haax poor habax cover, lid hahlo’o cloth, sail haks scold hakwhl gaff halayt Indian doctor, shaman hamom wrist hanak’ woman han’ii handle hanix to be thin hap cover hasak to want haseex rattle hasiyayks swallow (bird) hat’ marten hat’al’ cedar bark hats’ to bite haw’ don’t (negative imperative) haw’ to go home haw’ negative imperative verb hawaw mountain lion hawil arrow

208 Root Gloss haykw odor, spirit he to tell (trans) helt to be much, many het stand hets to send hum to be lonely hinda where, how, when hit’ scar, to heal hix fat, to be fat hlaa perfective aspect verb hlabix tired hlagok before OR in front of hlak to bend (trans.) hiak to be crooked hlalp to plane (wood) hiant to move hlap be deep hlaks nail, claw hlatsx fish tail hlaxs nail, claw hlbiy tell hleek to be worn hlgilkw to sweat hlguya jack-spring (small spring salmon) hlgwa’alt basket (for carrying water) hlimoo to help (trans) hlis finish hlit’ ball hliyun canned moosehide hlooç to be early hloots’ whitefish hloo squawfish hlo’oxs to kick hlok’ swallow hloks sun hloxs sun hlut’ax to boil ho again hoga to be correct, right, to resemble, be like anadromous fish, salmon (but also includes hon steelhead trout) hoobix spoon

209 Root Gloss hooni ankle hoo’oxs balsam tree hoox to use hukws be there (with someone), join hum’a lightning hup forehead hu’ums devil’s club huut to run away, flee (p1) huxws smoked salmon slices ihlak to be broken ihlee’e blood, to bleed ihleet’ red us necklace iiw’xt man (p1) in1 box (for storing food) is soapberry is to stink, to smell bad it call out ixsta to be sweet, tasty ixw to fish (with hook & line) jahi to eat up (trans) jahi to lose (intrans) jakw to kill jam to cook, boil jap to make, do (trans) jap small box trap japaan Japanese jayaas young beaver jayeehl steel trap jayn Chinese, Chinaman jilks to melt, thaw (intrans) jogo jaga across joiç to camp juxwt to untie k’ak’yotl’ ankle k’ats land, arrive k’eek to run away, flee’ k’eek drill k’eex tallow k’ibilap gravel, small rocks k’its’ee kidney k’uuhl year

210 Root iIIIIIIIIIIW G1o k’uukw’ tail kw’aat’a slug kw’adix jump around, flop kw’ast to be broken kwidats’ coat, jacket kw’oot to lose k’yo’o back k’yoots yesterday laagalt to look at laahl to be lying down laal bone-game, stick-game laam rum, alcoholic beverage laaxw trout laaxws lamp, lantern, light lakw fire, fuel lalt snake, worm lamk to be hot lan fish eggs laks to wash (one’sbody) laxs brush, pine needles lax’u above libast to sew libleet preacher, priest lidin to stand up (trans) (p1) ligil’ eyebrow li’lk feast lilp roll limx song, to sing lims grow (up) (p1) lip sew lis hang lit wedge lit wash (p1) litsx to count, read (trans) lix to watch loo to move in water, to swim (of fish) (p1) lo’op rock, stone loots’ elderberry bk to be rotten, old bk’ eel bk rotten, old lukw to move (change one’s residence)

211 Root Gloss luulak’ corpse, ghost luux alder tree luxw wart luxwt to refuse (to give) maak wash maas bark maay’ berries, fruit mahit to tell (intrans) m’al canoe m’al&a button malkw burn (up), put into fire m’alkw to put into fire (trans) m’alu to be crazy mamst to be funny, silly, foolish man to be left, be remaining (of food, etc.) mandi Monday mak put, place (trans) makt to leave (trans) m’as to grow up (p1) m’as to fart, to sting mats to arrive, run in number (of salmon) m’ats to hit, strike matx mountain goat ma’us to play m’as pants maxw burst out laughing (vi sg) mee pine cone meex to be sour (of milk, etc.) mehiat green/yellow mihl burn mihia gall, bile miin foot, base, boss mi’in smoke milikst crabapple milit steelhead trout milu dance m’isaax daylight miso’o sockeye salmon mit full m’it dusty, scattered miyup rice m’oja nearly

212 Root Gloss • momst to be foolish, nuts m’o’o pus moohi barrel, crock, a sort of fish-trap mooji almost mo’on salt m’ook to suck moos thumb, big toe moot cured, saved m’oot’ breast mo’ox meek, mild, quiet mukw sawbill duck m’ukw to catch (fish) mukw to be ripe mukw purple mumk’ to smile n’aa to complete an action naa who naasik’ raspberry n’akw to be long nak’ skirt nak’eeda muskrat nax snowshoe n’ax bait nda where, how, when nee no, to be none, negative verb neek hoof neek’ dorsal fin n’eexhl killerwhale, blackfish nimic skin on lip nis’in mink niso’o thimbleberry nis upper lip no’o hole no’ohl shoulder n’oo’o large birch-bark berry basket noosiic caterpillar n’usik newborn baby nuts’ snot nuuhlx to be wet (of a person) nuut dress up n’uw’ to die os dog

213 Root Gloss pdal ribs pdalt to climb pdeek tribe, phratry pdoohl siding phlo’on sea otter psa grey clay pt&aan totem pole kaap scratch k’aat cane k’aax wing, feather k’abaluu rifle kalaan’ behind, in back of k’alduux spoon (of alder wood or horn) ç’ap piece, part, relative ic’ap relative, kinsman kasx wild rice k’atx to be torn kelt top of hill k’esi knee k’esxk to be green, unripe k’ets’ chin k’ohl rope (of cedar) k’osx to spit up, spew ic’ots to cut ic’uhl to cut up (fish) (trans) k’uhlx chest sa day saak oolachen, candlefish sak’ to be stretched saks to leave (p1) saks to be clean sanakwa a caterpillar species sak to be sharp sak’ to split (intrans) saw’ns paper sdatxs stinging nettles sdil to accompany, go with sdim’oon humpback or pink salmon sdin to be heavy sduup stove se’e leg, foot seegal to be rough

214 Root Gloss sees spruce tree sga across the way sga herring sgan pitch, gum sgans elbow sgeks get hurt sgenx little finger sgi to be on sgi (gloss uncertain) sgimhl (gloss uncertain) sgolts to be ungrateful sgusiit potato sgwa to be square sgwat to joke sup to hurt siip’in to like (a person) siit vomit sil to be drunk, intoxicated sum set traps silkw middle, waist s’in bottom sins to be blind sint summer sip bone sip tough, hard sk’iik chickadee smaw’n maggot smax bear, meat sna thornberry so’o “doggy bag” sox frost sk’eex be dark sugwa sugar susiit potato suut to go for (in order to get) swan to blow sy’un glacier t’a louse t’aa to sit t’aahl to pick (berries) t’aap to hammer t’aboon padlock

215 Root Gloss t’ahlt to leave (trans) (p1) t’ahlt to put away, throw away (p1) t’ak to forget t’ak’ fold t’ak’aluts fox t’aks to dive t’akw lock t’asxu let’s (polite hortatory imperative) t’aw’ilt fish hook, safety pin t’ax lake t’eel’t to be fast t’eek to eat ravenously t’ikw’ navel, belly button t’ilix grease t’in fish weir, fishtrap t’in valley t’is to be big, large t’is to push, slug, push tk’i down tl’ook’ mud t’ook to eat t’oç to grab tk’a skin, hide ts’a’a eye, face ts’ahl to laugh ts’ak to go out, be extinguished (of a fire) ts’ak’ dish ts’al half-smoked salmon ts’amtx electricity, flash ts’ap the people of a village ts’ak nose ts’ak’ clam ts’ak’ dish ts’ee edge ts’eek to lick (trans) ts’eek make noise ts’eew’ interior, insides ts’eex to be full (after eating) ts’e juniper tree tsigins chicken ts’iihl snowbird ts’iikw to leak

216 1j Root JIItIIII!L Gloss ts’iip to close one’s eyes (intrans) ts’iip to tie ts’in to enter, go in ts’itxs waterfall ts’o’o to pause (intrans) ts’ook to stain (trans) ts’o’ot to skin (trans) ts’ok to be stuck, stick on (intrans) ts’ok’ salmon belly ts’uusx be small, little ts’uuts’ bird tun this tust that t’uuts’ charcoal txalt to put into fire (trans) (p1) txox halibut tyay’t thunder ubin to be pregnant ul drift umhl moss umk detest, dislike upji almost, just about uuic copper uut to bake (in ashes or oven) wa name w’a to find, to get to (someplace) w’aa’at scream waax paddle, under fin, to paddle w’alxa all wan deer wan to sit (p1) want to worry w’een fisher w’eek’s to find (trans) weex lynx wihi fir tree wilaax to know wilp house w’in tooth wineex food wis rain wisax sandbar

217 IhfII Root Gloss jijØ wist root wit collarbone, rack for drying salmon w’its’ to squeeze, mash (with hands) wo’os bowl, dish woot’ to sell wok to sleep wo’ to dig (trans) wo to bark (sing.) xa’a male slave xadaa moose at to be cold (of a person) xdeek to be dull xdii tea xeek foam, blossoms xhlaahl red willow bush hlgay sneak up hlgo’o to pay hlu to burst, explode, blow up xlilst to cough oo yawn xpaaw’ jaw xsaa only san to gamble xsan Skeena River xsdaa to win sgaak Eagle sgamks to steam (intrans) xsgoo to be first xshaneeks to snore xsihl lizard xsiip fine sand, sandbar xsink to disbelieve xsuuw’ edible hemlock bark ts’ay’ be thick (as a board) xwdaa mattress xwdakw to shoot (intrans) xwdax to be hungry xwsit fall, autumn xwts’aan totem pole ya’a spring salmon yahi smooth, slippery yahlx saliva, spittle

218 !!i Root JJ.:’ Gloss yal to lie, tell a lie yal to stir, beat yankw to be moldy y’ans leaf, grass, weeds y’ak to set (a snare), to be hanging yats to hit yaxw to hide (intrans) yee to go yeen cloud yeekt to share, give out yim to smell (trans) y’imk whiskers, beard yip land, ground yooiç eat yoot to roast, barbecue (over a fire) yo’oxs to wash face, hands, etc. yukw potlatch yuxw to follow

219