Measuring L2 Lexical Growth Using Hypernymic Relationships

Home , Lexicon, Order of acquisition, Polysemy, Synonym, Word sense

lang_508 langxml-als.cls (1994/07/13 v1.2u Standard LaTeX document class) 3-4-2009 :626

LANG lang_508 Dispatch: 3-4-2009 CE: N/A Journal MSP No. No. of pages: 28 PE: Matthew 1 Language Learning ISSN 0023-8333 2

3 4 5 Measuring L2 Lexical Growth Using 6 7 Hypernymic Relationships 8 9 Scott Crossley 10 Mississippi State University 11 12 Tom Salsbury 13 Washington State University 14 15 Danielle McNamara 16 University of Memphis 17 18 This study investigated second language (L2) lexical development in the spontaneous 19 speech of six adult, L2 English learners in a 1-year longitudinal study. One important 20 aspect of lexical development is lexical organization and depth of knowledge. Hyper- 21 nymic relations, the hierarchical relationships among related words that vary in relation 22 to their semantic specificity (e.g., Golden Retriever vs. dog vs. animal), are an important 23 indicator of both lexical organization and depth of knowledge. Thus, this study used 24 hypernymy values from the WordNet database and a lexical diversity measure to analyze 25 lexical development. Statistical analyses in this study indicated that both hypernymic 26 relations and lexical diversity in L2 learners increase over time. Additionally, lexical 27 diversity and hypernymic values correlated significantly, suggesting that as learners’ lex- 28 icons grow, learners have access to a wider range of hypernymy levels. These findings 29 are discussed in relation to developing abstractness in language, extending hypernymic knowledge, and the growth of lexical networks. 30 31 32 33 The research reported here was supported by the Institute of Education Sciences, US Department 34 of Education, through Grant IES R3056020018-02 to the University of Memphis. The opinions 35 expressed are those of the authors and do not represent views of the Institute or the US Department 36 of Education. The authors would also like to thank Dr. Philip McCarthy of the University of 37 Memphis for his statistical and methodological assistance. Additionally, the authors are indebted 38 to the three anonymous reviewers who provided critical assessments of the early and late versions of this article. 39 Correspondence concerning this article should be addressed to Scott A. Crossley, Department 40 of English, Mississippi State University, P.O. Box E, Starkville, MS 39759. Internet: sc544@ 41 msstate.edu

Language Learning 59:2, June 2009, pp. 307–334 307 C 2009 Language Learning Research Club, University of Michigan lang_508 langxml-als.cls (1994/07/13 v1.2u Standard LaTeX document class) 3-4-2009 :626

1 Crossley, Salsbury, and McNamara Hypernymic Relationships 2

3 Keywords hypernymy; hyponymy; superordinates; subordinates; lexical networks; cor- 4 pus linguistics; lexical diversity; computational linguistics 5 6 7 Introduction 8 Lexical proficiency in second language (L2) learners is far detached from 9 simplistic performance assessments such as number of words used or com- 10 prehended or the ability to match dictionary definitions with strings of letters. 11 It is now more fully recognized that an L2 learner’s lexical proficiency in- 12 cludes knowledge of syntactic properties, conceptual levels, sense relations, 13 and complex lexical association models (Haastrup & Henriksen, 2000; Huckin 14 & Coady, 1999). The lexical proficiency of L2 learners is of special interest 15 because misinterpretations of lexical items are key elements in global errors 16 that inhibit communication (de la Fuente, 2002; Ellis, 1995; Ellis, Tanaka, & 17 Yamakazi, 1994). Lexical proficiency is also crucial because the understanding 18 of lexical acquisition in relation to its deeper, cognitive functions can lead to 19 increased awareness of how learners process and produce an L2. Additionally, 20 studies into the development of L2 lexical proficiency, until recently, have been 21 an often neglected area of study in L2 acquisition (Meara, 2002). Those stud- 22 ies that have looked at lexical development have examined broad measures of 23 growth, such as lexical accuracy, lexical frequency, and lexical diversity (Polio, 24 2001). Studies that consider lexical growth in L2 learners are important be- 25 cause lexical growth strongly correlates with academic achievement (Daller, 26 van Hout, & Treffers-Daller, 2003). 27 This study explores one aspect of lexical proficiency: sense relations. Of 28 interest is the development of sense relations in L2 lexical acquisition and 29 how these sense relations relate to the richness of word knowledge and the 30 depth of vocabulary knowledge exhibited by L2 learners (Wesche & Paribakht, 31 1996). There are multiple variables that can be studied when analyzing depth 32 of L2 lexical knowledge. These include measures of polysemy, hypernymy 33 (superordinate and subordinate terms), synonymy, semantic overlap, and other 34 variables related to the conceptual meaning of words such as concreteness and 35 imageability. Although studies of L2 depth of lexical knowledge have been 36 rare, those studies conducted have generally demonstrated that L2 learners 37 generally do not have as much depth of lexical knowledge as do first language 38 (L1) speakers of those languages (Ordonez, Carlo, Snow, & McLaughlin, 2002; 39 Verhallen & Schoonen, 1993, 1998). 40 This study will concentrate on the growth of hypernymic relations 41 in L2 learners’ lexicon. Hypernymic relations are semantic links between

Language Learning 59:2, June 2009, pp. 307–334 308 lang_508 langxml-als.cls (1994/07/13 v1.2u Standard LaTeX document class) 3-4-2009 :626

1 Crossley, Salsbury, and McNamara Hypernymic Relationships 2

3 conceptually related words such as animal and dog. In this example, ani- 4 mal is semantically linked to dog but functions as a superordinate term, as it 5 is more abstract than the concrete, subordinate term dog. Although there have 6 been a few studies that have analyzed the hypernymic knowledge of bilingual 7 speakers (Defour & Kroll, 1995; Ordonez et al., 2002; Sharifian, 2002) and 8 L2 learners (Ijaz, 1986; Levenston & Blum, 1977), to our knowledge no study 9 has investigated the idea that as L2 learners acquire language; they also extend 10 and further develop lexical hypernymic relations. The possibility of investigat- 11 ing hypernymic relations in relationship to the growth of lexical networks was 12 raised by Haastrup and Henriksen (2000), but they did not directly investigate 13 hypernymy and instead measured synonymy, lexical gradation, and antonymy. 14 They argued that these were more crucial aspects of adjectives, which was the 15 linguistic category that was the focus of their study. This study, however, will 16 analyze verb and noun use and is thus well suited for investigating hypernymic 17 relationships. An analysis such as this could provide crucial information about 18 L2 lexical growth in three ways. First, it could provide evidence for the de- 19 velopment of lexical proficiency based on the use of extended sense relations. 20 Second, it could provide supporting evidence for the development of lexical net- 21 works in L2 language systems. Finally, it could provide evidence of abstraction 22 in L2 language use. 23 24 25 26 Hypernymy Defined 27 As stated earlier, this study will examine the growth of lexical hypernymic rela- 28 tions in L2 learners. It will not analyze the growth of L2 hypernymic concepts 29 because the data examined in this study come from adult L2 learners, who, 30 ostensibly, have fully developed conceptual knowledge of the world. Lexical 31 hypernymy is considered a fundamental semantic relationship that is founded 32 on the connection between general and specific lexical items (Chaffin & Glass, 33 1990; Haastrup & Henriksen, 2000). Hypernymic relations are hierarchical 34 associations between hypernyms (superordinate words) and hyponyms (subor- 35 dinate words). A hypernym is defined as a word that is more general than a 36 related word (animal compared to dog) and a hyponym is more specific than 37 a related word (dog as compared to animal). Another example of hypernymy 38 is the association between car and vehicle. In this case, car is the hyponym of 39 the hypernym vehicle because car has a narrower and more specific denotative 40 scope than vehicle, which would also include trucks, go-carts, golf carts, and 41 hearses.

309 Language Learning 59:2, June 2009, pp. 307–334 lang_508 langxml-als.cls (1994/07/13 v1.2u Standard LaTeX document class) 3-4-2009 :626

1 Crossley, Salsbury, and McNamara Hypernymic Relationships 2

3 An important aspect of hypernymy is the notion of basic level categories. 4 The theory of basic level categorization avers that an object at a specific level of 5 categorization has a superior status and is the concept that is used most often to 6 discuss an object (Brown, 1958). An important element that identifies whether 7 an object is a basic-level category is cue validity. Cue validity is premised 8 on the collection of features that distinguish an object from other objects. 9 The strongest cues are those that automatically place an object in a category 10 (e.g., gills for fish). Cue validity is important to basic categories because basic 11 categories are the level at which the largest number of attributes are contained. 12 All other members of a category (subordinates and superordinates) belong to 13 that category based on the number of features they share with the basic category. 14 Thus, car is a basic category, as it contains the most features that allow it to 15 be distinguished from other objects at a similar level (e.g., motorcycles, trucks, 16 and golf carts). Its subordinate, sedan, has lower cue validity because most of 17 its cues are shared across the category and few features distinguish it from a 18 car. In comparison, the superordinate category vehicle contains fewer shared 19 attributes of the basic category car (Rosch, Mervis, Gray, Johnson, & Boyes- 20 Braem, 1976). In summary, a basic category word is preferred for concept 21 labeling because subordinates are less distinctive but more informative, whereas 22 superordinates are more distinctive but not more informative. In contrast, basic- 23 level categories are associated with a large amount of information and they are 24 different from other categories at the same level. 25 There is much research that supports the notion of a natural preference to 26 use basic-level categories in speech and most research shows an overwhelming 27 propensity to describe the world using basic-level concepts (Murphy, 2004). In 28 studying the language habits of Tzeltal speakers, Berlin, Breedlove, and Raven 29 (1974) found that basic-level categorizations were named most readily, were 30 culturally more important, were more easily remembered, and were perceived 31 holistically. Downing (1980) reported that both native speakers of English and 32 Japanese had a tendency to categorize objects at the basic level compared to 33 using more concrete or abstract terms. Specifically, in the narratives collected by 34 Downing, she noted that native speakers of English used basic-level names 935 35 of the time and native speakers of Japanese used basic-level names 83% of the 36 time. Downing found that when superordinate or subordinate terms were used, 37 they were used for items that had poor basic-level coding, meaning the item 38 lacked prototypical features that fit the basic-level category (e.g., paddleball 39 or rooster). When subjects are asked to list properties for categories other 40 than basic-level categories, they are more likely to assign abstract properties 41 to superordinate terms and concrete properties to subordinate terms. Thus, the

Language Learning 59:2, June 2009, pp. 307–334 310 lang_508 langxml-als.cls (1994/07/13 v1.2u Standard LaTeX document class) 3-4-2009 :626

1 Crossley, Salsbury, and McNamara Hypernymic Relationships 2

3 higher on the hypernymic hierarchy a word is located, the more likely the word 4 will be abstract. The lower the word is located on the hierarchy, the more likely 5 the word will be concrete (Lakoff, 1987; Rosch, 1973; Tversky & Hemenway, 6 1984). 7 The competent use of hypernymic relations and abstraction is associated 8 with a speaker’saptitude to manage academic and formal registers (Snow, 1990; 9 Snow, Cancino, Gonzalez, & Shriberg, 1989; Snow, Cancino, De Temple, & 10 Schley, 1991). According to Ordonez et al. (2002), the more informal speech 11 is, the more hypernyms are used (i.e., “I took the animal there”), and the 12 more formal speech is, the more hyponyms are used (i.e., “I drove the Golden 13 Retriever to the doctor’s office”). Interest in academic and formal registers as 14 they relate to levels of abstraction has a long history in both linguistics and 15 cognitive science. For instance, Vygotsky (1962) argued that even minimal 16 academic writing skills require a high level of abstraction. 17 18 19 The Development and Knowledge of Hypernymy Relations 20 As noted by Ordonez et al. (2002), research on the acquisition of hypernymic 21 relations has demonstrated that hypernymic relations are more likely acquired as 22 the learners advance cognitively (Anglin, 1993; Snow, 1988; Vygotsky, 1962), 23 as they increase their levels of education (LeVine, 1980; Snow, 1990), as they 24 acquire more specific lexical knowledge (Wolter, 2001), and in an L1 rather 25 than an L2 (Meara, 1982; Soderman,¨ 1993). Most studies that have considered 26 hypernymic development have focused on L1 acquisition. A few have examined 27 bilingual conceptual development, not lexical development (e.g., Defour & 28 Kroll, 1995; Sharifian, 2002). It appears that no studies have analyzed the 29 development of hypernymic lexical relations in L2 learning, although some 30 studies have examined L2 learners’ use of hypernymy and the possible influence 31 of L1 conceptual categories on their use of L2 words. 32 Because it appears that no studies have looked at the development of hy- 33 pernymic lexical knowledge in L2 learners and few studies have analyzed L2 34 learners’ static, hypernymic, and lexical knowledge, a brief overview of hy- 35 pernymic studies in L1 learners will serve as an introduction to the idea. By 36 providing an overview of L1 hypernymic studies, we can establish a basic the- 37 oretical framework to serve as the foundation for this study and to guide our 38 eventual interpretations of the results. Seminal studies in the acquisition of hy- 39 pernymic conceptual knowledge in an L1 were conducted by Welch and Long 40 (1940) and Piaget (1972). These studies demonstrated that at a very young age 41 (2–6 years old), children can distinguish between hypernyms and hyponyms in

311 Language Learning 59:2, June 2009, pp. 307–334 lang_508 langxml-als.cls (1994/07/13 v1.2u Standard LaTeX document class) 3-4-2009 :626

1 Crossley, Salsbury, and McNamara Hypernymic Relationships 2

3 their L1. Additional studies that considered the acquisition of hypernymic rela- 4 tions in the L1 have demonstrated that children’sperception of categories differs 5 from that of adults and that children develop hierarchical patterns over time 6 (Anglin, 1977; Rosch, 1973; Rosch et al., 1976). More interestingly, children 7 display patterns of hypernymic, lexical, and conceptual development. Multiple 8 studies have affirmed that children first name and sort basic-level categories; 9 they then proceed toward superordinate categories (Berlin et al., 1974; Brown, 10 1958; Mervis & Crisafi, 1982; Murphy, 2004) and later toward subordinate 11 categories (Berlin et al.). This is likely because basic-level categories demon- 12 strate the most distinct level of action, contain names that are shortest and 13 most frequent, are the most natural level of categorization (Brown; Lakoff, 14 1987), and are the names most often found in the speech addressed to children 15 (Anglin). 16 When considering hypernymic categorizations and L2 learners, research 17 has supported the notion that L2 learners do not have the same access to hy- 18 pernymic relations as L1 speakers. Levenston and Blum (1977), for instance, 19 found that L2 learners use more words of general than of specific meanings. 20 This dependency on general terms leads L2 learners to make inappropriate 21 overgeneralizations and not adhere to expected uses of registers and colloca- 22 tions. In a study concerning the use of the prepositions on and over, Ijaz (1986) 23 found that L2 learners associated the prepositions most often with contexts that 24 were typical to their basic-level categorizations. Thus, L2 learners matched the 25 basic categorizations used by L1 speakers, but they failed to attribute similar 26 relationships to noncentral categories. This led to both the overuse and underuse 27 of central and noncentral conceptual categories. Specifically, on was used by 28 L2 learners in contexts in which more specific terms were required (upon and 29 onto). Overall, studies that consider the access L2 learners have to hypernymic 30 lexical categories demonstrate that L2 learners depend mostly on basic-level, 31 lexical categories. 32 One concern that is specific to L2 lexical acquisition is the role of cross- 33 linguistic influence and its function in semantic development. This is best 34 represented in Ijaz’s (1986) study in which she found that L2 learners tended 35 to transfer semantic relationships from their L1 system to their L2 system. 36 This transfer could influence the selection of contextually appropriate semantic 37 choices. Although Ijaz’s study was not directly related to hypernymic knowl- 38 edge, it is possible that cross-linguistic influence may exert certain effects on 39 learners’ understanding of how words are hierarchically related. However, Ijaz 40 did not address this possibility directly, and we contend that because hypernymic 41 relationships are relatively consistent across languages, there is no reason to

Language Learning 59:2, June 2009, pp. 307–334 312 lang_508 langxml-als.cls (1994/07/13 v1.2u Standard LaTeX document class) 3-4-2009 :626

1 Crossley, Salsbury, and McNamara Hypernymic Relationships 2

3 assume that L2 learners would process hypernymic structures differently be- 4 tween their L1 and L2. 5 6 7 Lexical Networks and Hypernymy 8 The acquisition of lexical items is now realized to be much more complex than 9 the simple memorization of a word and its definition. Even at the surface level, 10 word knowledge is more than simple definitions and includes the recognition 11 of sound patterns, orthography (Bogaards, 2001; Nation, 2005), and pronunci- 12 ation accuracy. In addition to these surface-level components, word knowledge 13 also includes syntactic specifications and links to conceptual content. It is 14 generally argued that the meaningful acquisition of lexical items involves the 15 use of inferencing strategies over memorization strategies and making connec- 16 tions between related words as opposed to memorizing definitions (Haastrup & 17 Henriksen, 2000; Huckin & Coady, 1999). The premise for such an approach 18 is founded on the notion of lexical networks, which is the idea that words 19 interrelate with other words to form clusters of words that act categorically. 20 These clusters connect to other clusters and other words, until entire lexicons 21 are developed based on interconnections (Ferrer i Cancho & Sole,´ 2001; Fer- 22 rer i Cancho, Sole,´ & Kohler,¨ 2004; Haastrup & Henriksen). One assumed 23 strength of these interconnections is that although words can have thousands of 24 related connections or nodes, the distance between each node (the connections 25 between words) is usually quite small (N. Ellis, in press). The relatively short Q1 26 distance between nodes provides the primary means for the quick growth of 27 lexical networks. Connections between words allow newly acquired words to 28 be easily assimilated within these networks because new words are not learned 29 in isolation, but through links to already learned words. As learners progress 30 lexically, they build lexical networks that are strengthened by differentiating 31 sense relations between words and within words (Haastrup & Henriksen). 32 Network building is also related to theories of emergentism, which is the 33 notion that language is a complex, adaptive system based on human interac- 34 tion. Importantly, language from an emergentist perspective is dynamic and 35 self-organizing (Ellis & Larsen-Freeman, 2006). Thus, lexical emergence is 36 thought to be based on the interaction of simple components (words) that 37 generate complex networks. According to Meara (2006), the emergent prop- 38 erties of a lexicon are similar to the properties of other types of complex 39 structures that exhibit behaviors that are not explicitly built into them. These 40 structures develop spontaneously through interactions between simpler com- 41 ponents. Specifically, this it taken to mean that L2 lexical networks emerge

313 Language Learning 59:2, June 2009, pp. 307–334 lang_508 langxml-als.cls (1994/07/13 v1.2u Standard LaTeX document class) 3-4-2009 :626

1 Crossley, Salsbury, and McNamara Hypernymic Relationships 2

3 from simple initial conditions that generate more complex networks from basic 4 lexical components. 5 An important approach to understanding lexical networks is complexity. 6 Most lexical relations are ordered by complexity in order to categorize and 7 identify elements of different relations (Herrmann & Chaffin, 1986; Klix & 8 van der Meer, 1980). In most network theories, hypernymy is considered a 9 primitive or simple sense relation. However, this does not mean that hypernymy 10 is not important. On the contrary, some theorists even suggest that hypernymy 11 is the single most important organizational system for lexical relations (Miller 12 & Teibel, 1991). This is because hypernymic relations allow for hierarchical 13 categorizations that define how hyponyms inherit properties from their related 14 hypernyms and allow set inclusion among category members. Hypernymy is 15 a foundational lexical relationship that is consistent with network models in 16 that it allows for the economical representation of lexical properties (Chaffin & 17 Glass, 1990; Murphy, 2004). Such properties allow for learners to generalize 18 about terms and allow for cognitive economy because every object is part of 19 a conceptual category, not its own conceptual category (Murphy). A sense 20 relation such as hypernymy that is based on an economy of representations 21 allows for the faster processing of language and ideas. This is compared to 22 other sense relations like synonymy, for which two concepts are treated as 23 equals, or polysemy, for which information is continuous (Chaffin & Glass). 24 Relations such as synonymy and polysemy are argued to be more difficult to 25 process. 26 27 28 Computationally Measuring Hypernymy 29 One method for measuring hypernymic development is through computational, 30 lexical databases meant to emulate lexical networks. Perhaps the best suited 31 database of this type is WordNet (Fellbaum, 1998; Miller, Beckwith, Fellbaum, 32 Gross, & Miller, 1990), which is a computational, lexical reference system 33 inspired by current psycholinguistic theories of lexical processing. In Word- 34 Net, English nouns, verbs, adjectives, and adverbs are organized into lexical 35 networks based on connections between related lexical concepts. In WordNet, 36 noun and verb hypernymy is measured through conceptual taxonomic hierar- 37 chies. For example, chair (a type of seat) has seven hypernym levels below 38 it: seat → furniture → furnishings → instrument → artifact → object → 39 entity. Thus, counterintuitively, a word located lower on the WordNet scale has 40 a higher hypernymy level and is thus more abstract. A word located higher on 41 the WordNet scale is more concrete.

Language Learning 59:2, June 2009, pp. 307–334 314 lang_508 langxml-als.cls (1994/07/13 v1.2u Standard LaTeX document class) 3-4-2009 :626

1 Crossley, Salsbury, and McNamara Hypernymic Relationships 2

3 WordNet hypernymy values have been used in various studies. Most of 4 these studies involve disambiguating word sense and semantic relations for 5 natural language processing tasks (Fragos & Maistros, 2005; Resnik, 1992), 6 investigating semantic similarity (Leacock & Chodorow, 1998), and measuring 7 conceptual density (Agirre & Rigaus, 1996). In L2 studies, WordNethypernymy 8 values have been used to compare simpliﬁed and authentic reading text for dif- 9 ferences in levels of abstraction (Crossley, Louwerse, McCarthy, & McNamara, 10 2007; Crossley & McNamara, in press). This study will use WordNet values Q2 11 as a method to track the range of hypernymic levels used by L2 learners over 12 time and to show how the range of meanings used by L2 learners develops. 13 By proxy, this development can be used to examine the growth of L2 lexical 14 networks. 15 16 17 Method 18 Our purpose in this study is to explore whether L2 learners develop hypernymic 19 lexical relationships as their lexicon grows and, if so, how such growth relates to 20 the development of lexical networks. To accomplish this, we use the WordNet 21 database to assess the spoken utterances of L2 learners using longitudinal data. 22 We examine whether WordNet indexes of hypernymic relations increase or 23 decrease as learners acquire an L2. We make the prediction that as L2 learners 24 develop lexical knowledge, there will be a corresponding increase in lexical 25 hypernymy levels in a manner similar to the growth of hypernymic relations 26 in an L1. Findings such as these would suggest that as L2 learners’ lexicons 27 grow, the L2 learners begin to make simple, hierarchical hypernymic lexical 28 connections. This result would provide support for the development of abstract 29 language ability as well as for the development of lexical networks in L2 30 learners. 31 Because of the slow, developmental nature of lexical networks (Haastrup 32 & Henriksen, 2000), we selected a longitudinal approach to data collection and 33 analysis. A longitudinal approach is better suited for analyzing the building 34 of lexical networks because it allows for the investigation of language learners 35 over a long period of time. In contrast, a cross-sectional study examines separate 36 groups of participants, each at different developmental states or levels. This 37 type of approach does not allow for the examination of gradual increases in 38 lexical proﬁciency. Thus, the longitudinal approach is better suited to capture 39 gradual lexical developments, such as lexical networks. However, it is more 40 challenging to obtain large numbers of participants in longitudinal studies. 41 As such, the statistical power of longitudinal studies emerges from having a

315 Language Learning 59:2, June 2009, pp. 307–334 lang_508 langxml-als.cls (1994/07/13 v1.2u Standard LaTeX document class) 3-4-2009 :626

1 Crossley, Salsbury, and McNamara Hypernymic Relationships 2

3 greater number of observations per participant rather than large numbers of 4 participants. 5 6 Participant Selection 7 The participants used in this study were a group of L2 English learners enrolled 8 in an intensive English program at a large American university. They were 9 interviewed every 2 weeks (not including program and university breaks) over a 10 1-year period. The elicitation sessions were scheduled such that they coincided 11 with the students’ regular speaking class. Learners’ proﬁciency levels were 12 tested upon arrival to the program, and all participants in the study tested into the 13 lowest proﬁciency level, Level 1, of a six-level program. The learners continued 14 their participation in the study as they progressed through the program. 15 The focus of this study is on six of the learners in the original cohort of 16 50 students. The other 44 learners were dropped from the analysis because of 17 large gaps in the elicitation data during the year of observation or because they 18 did not complete the year. Each learner in the study was given a pseudonym; 19 this article reports on data from Marta (Spanish L1), Takako (Japanese L1), 20 Eun Hui (Korean L1), Faisal (Arabic L1), Kamal (Arabic L1), and Jalil (Arabic 21 L1). The participants ranged in age from 18 to 29 years and had all successfully 22 completed high school in their country of origin. None of the learners had 23 lived in the United States for more than 3 weeks prior to the start of the study. 24 All learners had studied English in their native secondary schools. Only Marta 25 reported using English in a professional setting, but because of her limited 26 ability with English, she was sponsored by her company to study English 27 intensively for 1 year. 28 Interviewers were recruited from a graduate-level course in second language 29 acquisition (SLA) taught at the university. Participants (interviewers and L2 30 learners) were given a variety of elicitation materials, including emotion cards 31 (Rintell, 1989), picture description tasks, questions, and topics for discussion 32 as prompts to promote spontaneous speech. The participants were also free 33 to introduce their own spontaneous topics into the conversation. Language 34 data were collected in naturalistic settings; that is, although interviewers came 35 to the sessions prepared with various topics from which the learners could 36 choose, the sessions were characterized by naturally occurring discourse. In 37 some cases, when the scheduled elicitation session contained more learners 38 than interviewers, learners were paired with an interviewer, providing discourse 39 data between the L2 English learner and his or her native-speaking interviewer 40 as well as learner-to-learner data. Elicitation sessions generally lasted between 41 30 and 45 min. The sessions were tape-recorded and later transcribed. Because

Language Learning 59:2, June 2009, pp. 307–334 316 lang_508 langxml-als.cls (1994/07/13 v1.2u Standard LaTeX document class) 3-4-2009 :626

1 Crossley, Salsbury, and McNamara Hypernymic Relationships 2

3 Table 1 Descriptive statistics for longitudinal language data 4 Number of Average number Average Average number 5 meetings/ of words number of of words 6 Learner transcripts per utterance utterances per text 7 8 Eun Hui 18 21.31 52.17 1,120.67 9 Faisal 13 33.42 71.08 1,870.07 Takako 18 19.40 51.00 1,470.72 10 Kamal 15 23.75 50.27 1,216.20 11 Jalil 17 38.82 61.76 2,359.77 12 Marta 18 33.31 63.61 1,912.00 13 14 15 the focus of the current article is on sense relations, we sought longitudinal data 16 from multiple learners engaged in discourse on various topics. In comparison 17 to studies using data from a limited number of oral or written topics, the method 18 used in the current study elicited a greater quantity and variety of vocabulary. 19 It also captured learners in a naturalistic environment and is arguably a better 20 representation of the learners’ oral proficiency. This method of longitudinal 21 data collection has also been successfully employed in research on the SLA of 22 tense-aspect relations (Bardovi-Harlig, 2000). 23 24 Corpus 25 The spoken data collected from the six learners was transcribed and forms the 26 basis for this analysis. A total of 99 transcripts were collected. Descriptive data 27 for the corpora of each learner are presented in Table 1. In preparation for the 28 analysis of the learner corpus, transcriptions of each elicitation session were 29 modified in the following ways: Interjections such as ah, uhm, and yea were 30 deleted, as were any words that were clearly non-English words, such as a word 31 in the learner’s native language or an invented word. All punctuation except 32 the period and question mark was eliminated from the transcriptions. Each 33 elicitation session was saved as a single text file containing the oral production 34 of only the learner in focus, not the interviewer or other learners participating in 35 the session. The text file was manually and electronically checked for spelling 36 errors. 37 38 Coh-Metrix 39 The computational tool Coh-Metrix (Graesser, McNamara, Louwerse, & Cai, 40 2004) was used to assess the lexical development of the learners in this study 41 (general lexical growth and hypernymy). Coh-Metrix measures cohesion and

317 Language Learning 59:2, June 2009, pp. 307–334 lang_508 langxml-als.cls (1994/07/13 v1.2u Standard LaTeX document class) 3-4-2009 :626

1 Crossley, Salsbury, and McNamara Hypernymic Relationships 2

3 text difficulty at various levels of language, discourse, and conceptual analysis. 4 The system integrates lexicons, pattern classifiers, part-of-speech taggers, syn- 5 tactic parsers, shallow semantic interpreters, and other components that have 6 been developed in the field of computational linguistics (Jurafsky & Martin, 7 2000). 8 Coh-Metrix also provides WordNet measures for hypernymy. In Coh- 9 Metrix, texts are first processed through the Brill Tagger (Brill, 1995), which 10 assigns part of speech tags to all words. Coh-Metrix then measures the hyper- 11 nymy values for both nouns and verbs using WordNet. WordNet hypernymy 12 values are provided on a normalized scale, with 1 being the highest hypernym 13 score and all related hyponym values decreasing after that. As noted earlier, 14 this means that a lower score reflects an overall use of more abstract words, 15 whereas a higher score reflects an overall use of more concrete words. Thus, 16 entity, as a possible hypernym for the noun chair, would be assigned the num- 17 ber 1. All other possible hyponyms of entity as it relates to the concept of a 18 chair (e.g., object, furniture, seat, chair, camp chair, and folding chair) would 19 receive higher values. Similar values are assigned for verbs (e.g., hightail, run, 20 and travel). This is a product of a categorical structure that requires the most 21 abstract concept to be the most stable concept, as it will be shared by all lexical 22 items. This means that all lexical items will share a default value of 1 (an 23 abstract relation), but not all lexical items will share larger values (concrete 24 relations). In this study, we use these WordNet values to evaluate the growth of 25 hypernymic relations in L2 learners. 26 In addition to WordNet values, Coh-Metrix also reports word meaning val- 27 ues from the MRC Psycholinguistic Database (Coltheart, 1981). To provide 28 support for the hypothesis that L2 hypernymic relations increase with time and 29 lexical growth, we will also examine MRC concreteness values. This is because 30 we hypothesize that as hypernymic relations in L2 learners’ oral discourse in- 31 crease (with lower WordNet hypernymy values), their lexicons will become 32 more abstract. To test this hypothesis, we will compare WordNet hypernymy 33 values to MRC concreteness values, with the expectation that MRC concrete- 34 ness values will decrease with WordNet hypernymy values. MRC concreteness 35 values are based on the work of Paivio, Yuille, and Madigan (1968), Toglia and 36 Battig (1978), and Gilhooly and Logie (1980), who asked participants to score 37 nouns based on concreteness on a numerical scale (from 1 to 7). A word that 38 refers to an object, material, or person generally receives a higher concreteness 39 score than an abstract word (Toglia & Battig). 40 To examine L2 lexical growth, we will use the Measure of Textual 41 and Lexical Diversity (MTLD: McCarthy, 2005) that is also reported by

Language Learning 59:2, June 2009, pp. 307–334 318 lang_508 langxml-als.cls (1994/07/13 v1.2u Standard LaTeX document class) 3-4-2009 :626

1 Crossley, Salsbury, and McNamara Hypernymic Relationships 2

3 Coh-Metrix. MTLD is a lexical diversity (LD) measure. MTLD overcomes 4 potential confounds of text length by separating a text into segments that have 5 a TTR value of .71. The total segments in the text are then divided by the Q3 6 total number of words in the text (the forward value). The whole process is 7 then repeated starting at the end of the text rather than the beginning (the re- 8 verse value). The forward and reverse values are then summed and divided by 9 2 for the MTLD value of the text. MTLD was selected to link the WordNet 10 hypernymy data to lexical growth because LD measures are commonly used to 11 analyze lexical growth (Jarvis, 2002; Malvern & Richards, 2002; Polio, 2001). 12 MTLD is also similar to other measures of LD such as D (Malvern, Richards, 13 Chipere, & Duran, 2004) or TTR (Miller, 1981; Templin, 1957), but, unlike 14 other LD measures, MTLD does not vary as a function of text length for text 15 segments whose length is in the 100–2,000-word range. MTLD also allows for 16 comparisons between text segments of largely different lengths (up to 2,000 17 words) and produces reliable results over a wide range of genres while strongly 18 correlating with other LD measures (McCarthy). Thus, MTLD is able to as- 19 sess differences of lexical diversity between different texts (both spoken and 20 written), even when those texts may be considerably different in terms of text 21 length. We use MTLD in this study to assess if L2 learners exhibit signs of 22 lexical growth over the course of a year. 23 24 25 Results 26 To test the hypothesis that time spent learning English relates to the development 27 of hypernymic relations, a repeated-measures analysis of variance (ANOVA) 28 was conducted using the WordNet results from Coh-Metrix. The ANOVA was 29 used to track the linear trend of the WordNet values over the increasing tem- 30 poral intervals and to test the assumption that as time spent learning English 31 increased, WordNet hypernymic values would decrease, signaling an increase 32 in hypernymic relations. Because participants did not share all the same tem- 33 poral data points, the ANOVA test analyzed development on a trimester basis. 34 This allowed for breaks in the data related to winter and spring recesses to be 35 considered as well as missing data points resulting from participant absences. 36 Because data were available for the 2nd and 4th weeks and the 50th and 52nd 37 weeks for all six learners, they were included. These data points were analyzed 38 with data from the 16th week and the 32nd week as well. This ANOVA was 39 supplemented with pairwise comparisons to identify signiﬁcant differences 40 within the temporal progression. Because not all participants shared all tempo- 41 ral points, the repeated-measures ANOVA analyzed an unequal sample of the

319 Language Learning 59:2, June 2009, pp. 307–334 lang_508 langxml-als.cls (1994/07/13 v1.2u Standard LaTeX document class) 3-4-2009 :626

1 Crossley, Salsbury, and McNamara Hypernymic Relationships 2

3 temporal points available. Thus, a separate growth modeling analysis (a linear 4 curve estimation) to control for random effects was conducted. This allowed for 5 both shared temporal data points (N = 36) and all temporal data points (N = 6 99) from the L2 learners to be analyzed and examined if the linear trend reported 7 in the ANOVA was supported individually and collectively. 8 To support the ANOVA and linear curve estimation findings, two Pearson’s 9 product-moment correlation tests were also conducted. The first was the cor- 10 relation test between time spent learning English and the WordNet hypernymy 11 values using all available data points (N = 99). For this analysis, we predicted 12 negative correlations because WordNet hypernymy values would actually de- 13 crease as time increased, signaling a wider range of hypernymic knowledge. 14 A second correlation was conducted between WordNet hypernymy values and 15 MRC concreteness values for all data points (N = 99). For this analysis, we 16 predicted a positive correlation because MRC concreteness values should de- 17 crease with an increase in abstract words, which would correlate to a decrease 18 in WordNet hypernymy values as hypernymic knowledge develops. 19 Table 2 shows that the L2 learners’ WordNet hypernymy values decreased 20 as a function of time, defined as the 2nd, 4th, 16th, 32nd, 50th, and 52nd weeks 21 of learning, F(5, 25) = 13.57, p < .001, η2 = .73. Pairwise comparisons of 22 the WordNet hypernymy values from the 2nd week demonstrated significant 23 differences from other temporal data points beginning in the 16th week and 24 continuing until the 52nd week (see Table 3 for details). There was also a 25 significant linear trend, F(1, 25) = 19.31, p < .01, η2 = .78. The linear curve 26 estimate conducted to support the ANOVA findings demonstrated a significant 27 linear trend for all participants, R2 = .87, F(1, 4) = 25.62, p < .01. Indi- 28 vidual linear curve estimates demonstrated significant linear trends in five of 29 the six participants in the study (see Table 4). These findings indicate that as 30 time spent learning English increased, hypernymy levels increased. These find- 31 ings also provide evidence that hypernymic relations develop with time spent 32 33 Table 2 Mean and standard deviations for WordNet hypernymy values 34 35 Week Mean SD 36 2 1.44 0.22 37 4 1.34 0.18 38 16 1.23 0.15 39 32 1.19 0.13 40 50 1.02 0.08 41 52 1.05 0.11

Language Learning 59:2, June 2009, pp. 307–334 320 lang_508 langxml-als.cls (1994/07/13 v1.2u Standard LaTeX document class) 3-4-2009 :626

1 Crossley, Salsbury, and McNamara Hypernymic Relationships 2

3 ∗ ∗ ∗ 4 5

6 0.04 (0.02) 0.38 (0.09) 0.28 (0.08) 0.17 (0.06) 0.04 (0.04) = = = 7 = =− 8 Diff Diff Diff Diff 9 10 11 ∗ ∗ ∗ 12 13 0.08 (0.04) Diff 14 0.42 (0.08) 0.32 (0.08) 0.21 (0.06) = = = = 15 Diff 16 Diff Diff Diff 17

18 ∗ ∗ ∗ 19 20 0.13 (0.04) 21 0.34 (0.06) 0.24 (0.07) = 22 = =

23 Diff Diff 24 25 ∗ ∗ 26 27 28 0.11 (0.04) 0.21 (0.04) = 29 = 30 31 32 33 34 35 36 0.10 (0.05) Diff = 37 38

39 Pairwise comparison of temporal differences in WordNet hypernymy values Diff denotes the average difference between cosines (standard error). 40 .05. < 24 Diff Diff p Week416325052 1632 Diff 41 50 Table 3 Note. ∗

321 Language Learning 59:2, June 2009, pp. 307–334 lang_508 langxml-als.cls (1994/07/13 v1.2u Standard LaTeX document class) 3-4-2009 :626

1 Crossley, Salsbury, and McNamara Hypernymic Relationships 2

3 Table 4 WordNet hypernymy linear curve estimations for individual participants 4 L2 learner NR2 Fp 5 6 Eun Hui 18 0.67 32.69 <.001 7 Faisal 13 0.79 40.77 <.001 < 8 Takako 18 0.76 51.75 .001 > 9 Kamal 15 0.04 0.53 .05 Jalil 17 0.74 42.19 <.001 10 Marta 18 0.75 48.28 <.001 11 12 13 14 learning English. In addition, a correlation between the L2 learners’ time spent 15 studying English and their WordNet hypernymy values also demonstrated sig- 16 nificant negative correlations (r =−0.65, p < .001, N = 99). As expected, a 17 correlation between the L2 learners’ WordNet hypernymy values and their MRC 18 concreteness values showed a significant positive correlation (r = 0.62, p < 19 .001, N = 99). These findings help to demonstrate that as L2 learners’ time 20 studying English increases, hypernymic relations develop and their lexicon 21 becomes more abstract. 22 To test the assumption that learners’ lexical diversity grows as time spent 23 learning English increases, a repeated-measures ANOVA was conducted using 24 the MTLD results from Coh-Metrix. In a fashion similar to the hypernymy 25 analysis, this ANOVA was used to track the linear trend of the MTLD values 26 over the increasing temporal intervals. Like the hypernymy data, participants 27 did not share all the same temporal data points, so the ANOVA analyzed devel- 28 opment on a trimester basis, but it included the 2nd and 4th weeks and the 50th 29 and 52nd weeks. As in the hypernymy analysis, this ANOVA was also sup- 30 plemented with pairwise comparisons to identify significant differences within 31 the temporal progression and polynomial contrasts to test for linear trends in 32 the data. A separate growth modeling analysis (a linear curve estimation) was 33 also conduced to support whether the linear curve reported in the ANVOA 34 was supported individually and collectively. To support the ANOVA and linear 35 curve estimation, the Pearson’s product-moment correlation test between time 36 spent learning English and MTLD values using all available data points (N = 37 99) was also calculated. For this analysis, we predicted a positive correlation 38 because lexical diversity should increase as a function of time. To examine the 39 link between the development of lexical diversity and hypernymic relations, a 40 correlation between the MTLD values and the WordNet hypernymy values for 41 all data points (N = 99) was conducted. In this analysis, we predicted a negative

Language Learning 59:2, June 2009, pp. 307–334 322 lang_508 langxml-als.cls (1994/07/13 v1.2u Standard LaTeX document class) 3-4-2009 :626

1 Crossley, Salsbury, and McNamara Hypernymic Relationships 2

3 Table 5 Mean and standard deviations for MTLD values 4 Week Mean SD 5 6 2 28.43 7.27 7 4 25.37 4.55 16 32.26 7.27 8 32 31.12 3.78 9 50 34.88 4.25 10 52 35.43 2.92 11 12 13 correlation because MTLD values were expected to increase as a function of 14 time, whereas WordNet hypernymy values were expected to decrease. 15 Table 5 shows that the L2 learners’ MTLD values increase as a function 16 of time, defined as the 2nd, 4th, 16th, 32nd, 50th, and 52nd weeks of learning, 17 F(5, 25) = 7.41, p < .001, η2 = .60. Pairwise comparisons of the MTLD val- 18 ues demonstrated significant differences between the 2nd week and the 52nd 19 week (see Table 6 for more details). There was also a significant linear trend, 20 F(1, 25) = 22.83, p < .01, η2 = .64. The linear curve estimate conducted 21 to support the ANOVA findings demonstrated a significant linear trend for all 22 participants, R2 = .92, F(1, 4) = 45.55, p < .01. Individual linear curve esti- 23 mates demonstrated significant linear trends in five of the six participants in the 24 study (see Table 7). These findings indicated that as time spent learning English 25 increased, MTLD values increased. As predicted, the correlation between the 26 L2 learners’ time spent learning English and their MTLD values demonstrated 27 a significant positive correlation (r = 0.45, p < .001, N = 99). In addition, the 28 correlation between the L2 learners’ MTLD values and their WordNet hyper- 29 nymy values showed a significant negative correlation (r =−0.40, p < .001, 30 N = 99) as predicted. These findings provide evidence that as learners spend 31 time studying the English language, their lexical diversity increases. Addition- 32 ally, it provides evidence that the development of hypernymic relations in L2 33 learners is related to growth in lexical diversity. 34 35 36 Discussion 37 This analysis has demonstrated that as time studying an L2 increases, there 38 is a corresponding increase in the range of hypernymy levels available to L2 39 learners. It has also demonstrated that hypernymy growth is related to L2 lex- 40 ical growth. These findings have important implications for the development 41 of abstract and formal language use in L2 learners as well as the growth of L2

323 Language Learning 59:2, June 2009, pp. 307–334 lang_508 langxml-als.cls (1994/07/13 v1.2u Standard LaTeX document class) 3-4-2009 :626

1 Crossley, Salsbury, and McNamara Hypernymic Relationships 2

3 ∗ ∗ ∗ ∗ 4 5 6 10.05 (1.40) 7.23 (2.34) 7.05 (1.81) 6.99 (2.48)

7 –.55 (1.80) = =− =− =− 8 =− Diff 9 Diff Diff Diff 10

11 ∗ ∗ ∗ 12 13 6.51 (1.71) 9.50 (1.10) 6.68 (2.11) 14 6.44 (3.07) Diff =− =− =− 15 =− 16 Diff Diff 17 18 ∗ 19 20 21 0.17 (1.48) Diff 2.99 (0.72)

22 0.07 (2.85) Diff =− =− 23 = 24 25 26 27 28

29 2.82 (1.43) Diff .24 (3.05) Diff =− 30 = 31 32 33 34 35

36 3.06 (2.78) Diff 37 = 38

39 Pairwise comparison of temporal differences in MTLD values Diff denotes the average difference between cosines (standard error). 40 .05. < 24 Diff Diff p

41 Week1632 450 16 32 Diff 50 52 Table 6 Note. ∗

Language Learning 59:2, June 2009, pp. 307–334 324 lang_508 langxml-als.cls (1994/07/13 v1.2u Standard LaTeX document class) 3-4-2009 :626

1 Crossley, Salsbury, and McNamara Hypernymic Relationships 2

3 Table 7 MTLD linear curve estimations for individual participants 4 L2 learner NR2 Fp 5 6 Eun Hui 18 0.24 4.99 <.05 7 Faisal 13 0.32 5.05 <.05 < 8 Takako 18 0.41 10.88 .01 < 9 Kamal 15 0.5 12.10 .01 Jalil 17 0.05 0.87 >.05 10 Marta 18 0.67 32.53 <.001 11 12 13 14 lexical networks and L2 lexical proficiency. The data presented here provide 15 evidence that L2 learners make rapid and significant changes in their use of 16 hypernymic terms over the course of 1 year. Spoken L2 vocabulary exhibits 17 significantly higher levels of hypernymy at 16 weeks, 32 weeks, 50 weeks, and 18 52 weeks than during lexical production in the second week. Significant differ- 19 ences between all weeks up until the 50th week were also found. These findings 20 not only show that learners begin to use more abstract language as their lexicon 21 develops, but they also point toward the growth of simple, hierarchical lexical 22 connections. This finding suggests the development of L2 lexical networks. 23 These findings should not be viewed holistically as a strict linear progression 24 from the concrete to the abstract, but, more importantly, as the development of a 25 range of available lexical choices. Thus, these data should not be solely viewed 26 as allowing a learner to move from the term dog to entity, but, if the adopted 27 framework is accurate, also for predicting that the learner would have access 28 to the lexical levels that fall between dog and entity, such as mammal, animal, 29 organism, and object. Exemplified in this way, the lexical acquisition of L2 30 learners is not a simple migration from concrete to abstract, but it a collection 31 of steps that ultimately provides L2 learners with not only the ability to use 32 abstract language but, more importantly, also a greater range of available lexical 33 items. The mean hypernymy values support such a movement (see Table 2); 34 however, the results are speculative and it will be the task for future studies to 35 identify the degree to which learners have access to intermediate hypernymic 36 levels. Interestingly, though, as time spent learning English increases, the stan- 37 dard deviations between learners decreases. This demonstrates that learners’ 38 hypernymic values converge as their vocabularies develop and become more 39 abstract, likely signaling the lack of choice available at the top end of the hier- 40 archy (the most abstract hypernyms). It could also signify that learners are not 41 taking advantage of the range of lexical choices available to them.

325 Language Learning 59:2, June 2009, pp. 307–334 lang_508 langxml-als.cls (1994/07/13 v1.2u Standard LaTeX document class) 3-4-2009 :626

1 Crossley, Salsbury, and McNamara Hypernymic Relationships 2

3 From a practical standpoint, the WordNet as well as the MRC concreteness 4 findings do support the idea that L2 learners begin to use more abstract language 5 and this is also important. Abstract language use is not only fundamental to 6 most forms of communication, but it is also important for specific language 7 genres such as professional and academic genres. Abstract language use also 8 demonstrates that L2 learners likely become more adept at using formalized 9 speech (Ordonez et al., 2002). Additionally, advancing levels of hypernymic 10 relations are related to the development of specific lexical knowledge (Wolter, 11 2001). This growth in abstract language should allow the L2 learner to speak 12 about ideas that are displaced from the “here and now” and should allow L2 13 learners to discuss broad concepts and decontextualized ideals. Abstraction 14 likely gives the L2 learner the opportunity to discuss ideas in general terms, 15 instead of discussing specific objects. 16 Examples selected from the data supporting a movement toward more ab- 17 stract language use are provided below. These examples are not representative; 18 they are meant to be illustrative. The first examples are taken from Faisal’sdata. 19 In (1), taken from the second week, Faisal is discussing arranged marriages. 20 In (2), taken from the 50th week, Faisal is discussing his English ability. In 21 (1), there are multiple examples of specific language use (restaurant, eat, sit, 22 salt, cup, clothes, mother, and sister) and few examples of general terms (home 23 and woman). In (2), the lexical choices are much more generalized ( feel, man, 24 language, understand, opinion, someone, and something). 25 (1) Some day I go to a restaurant, I eat here, I sit here eat. He’s want salt. 26 I have salt. He say, cup, clothes. Come my home. He’s come my home, 27 and my mother, no my mother. Do you know my mother’s sister? No my 28 mother’s sister. She’s woman. what name? 29 30 (2) I feel like a new man, because I can speak new language. I’m someone 31 different. I can another language. I can understand another opinion, not, 32 Arabian like, American opinion, so American something, you know. You 33 learn something you feel something different, you know, that’s me. 34 Excerpts selected from Jalil’sdata also demonstrate that there is a movement 35 from the concrete to the abstract. In (3), taken from the third week, Jalil is talking 36 about statues of famous people. His utterances contain multiple examples of 37 more concrete lexical items such as newspaper, month, tourist, stone, John 38 Kennedy, and picture and only a few abstract terms. Example (4) taken from 39 the 50th week demonstrates the use of more hypernymic language use. In 40 this example Jalil is discussing his future. He uses multiple examples of more 41 abstract lexical choices such as fantasy, future, something, and universe.

Language Learning 59:2, June 2009, pp. 307–334 326 lang_508 langxml-als.cls (1994/07/13 v1.2u Standard LaTeX document class) 3-4-2009 :626

1 Crossley, Salsbury, and McNamara Hypernymic Relationships 2

3 (3) I read in one newspaper last, uh, one month. The visitor of tourist see one, 4 the stone, one stone like, like the, uh, John Kennedy. The same look like. 5 He get the, uh, picture. 6 (4) But I have a lot of fantasy in the future, like working in NASA company, 7 ah, like discovering something new, nobody know about it, universe, or 8 something in astronomy. So, this is one of my most fantasy that I looking 9 to. 10 11 Examples (3) and (4) were extracted from Eun Hui’sdata. In (5), taken from 12 the third week, Eun Hui is discussing loneliness. In doing so, she depends on 13 very concrete lexical items such as Thursday, evening, raining, parents, room, 14 and night. Later, in the 50th week, Eun Hui contrasts American and Korean 15 classrooms. During this discussion, Eun Hui uses more abstract terms such as 16 style, remember, things, attention, and person. 17 (5) Thursday evening, raining. Me very very lonely. I think my country, my 18 parents, my family, very very sad. And my room, my room, night, only, 19 only alone, lonely. 20 21 (6) But this style is, maybe, that’swrong, because we have to remember many, 22 many things, during our class. Just we have to pay attention to the teacher. 23 If, if the teacher asked, ask a student, just, ah, indicated one person, the 24 person have to answer. But depend on person. 25 The statistical findings also suggest the growth of lexical networks in L2 26 learners. Although hypernymy is a simple sense relation, it is important because 27 it allows for hierarchical categorization. Evidence that shows the growth of 28 hierarchical ordering systems might also demonstrate that learners are able 29 to differentiate lexical sense relations as they progress and thus are likely in 30 the process of constructing lexical networks (Haastrup & Henriksen, 2000). 31 Hypernymic relations, as a part of lexical network building, are important 32 because they allow for lexical properties to be represented economically in 33 hierarchical fashion. This approach to lexical ordering should allow for related 34 words to be stored in categorical clusters. This would permit words to be 35 processed and recalled more quickly and could affect the comprehensibility 36 of the L2 learners’ output. Connections between related words also provide 37 learners with an efficient system from which to associate newly encountered 38 words with already acquired words. These links should help L2 learners more 39 quickly acquire lexical proficiency (R. Ellis, 1994; Haastrup & Henriksen) 40 because they would allow learners to draw inferences about lexical hierarchies 41 as well as use taxonomical reasoning to make generalizations about words

327 Language Learning 59:2, June 2009, pp. 307–334 lang_508 langxml-als.cls (1994/07/13 v1.2u Standard LaTeX document class) 3-4-2009 :626

1 Crossley, Salsbury, and McNamara Hypernymic Relationships 2

3 not stored in memory (Murphy, 2004). Such rapid growth can be observed 4 in the data presented in this study, especially when considering the finding 5 that significant hypernymic growth was realized within the first 4 months of 6 learning English in a native environment as well as significant increases across 7 all temporal intervals in the first 8 months. Findings such as these lend credence 8 to notions of lexical networks as well as theories of language emergence. 9 Another interesting finding of this study is that adult L2 learners appear to 10 follow r lexical learning patterns similar to those followed by L1 learners. These 11 patterns can be detected in the changes found in the mean WordNet hypernymy 12 values for all the learners (see Table 2). These mean values demonstrate that 13 L2 learners move from more concrete terms to more abstract terms in a manner 14 similar to L1 learners (Berlin et al., 1974; Brown, 1958; Mervis & Crisafi, 15 1982; Murphy, 2004). Because each word in WordNet will produce different 16 hypernymy values, it is difficult to judge whether or not the initial words used 17 by the L2 learners in this study are generally basic category words, but this is 18 likely the case considering past studies which show a propensity for L2 learners 19 to depend on basic category terms (Ijaz, 1986; Levenston & Blum, 1977). 20 Although speculative, the movement toward more subordinate terms found in 21 the 52nd week may correspond to the last stage of word learning found in 22 L1 learners: the movement back toward more specific referents (Berlin et al., 23 1974). 24 25 26 Conclusions 27 This study has demonstrated that lexical hypernymic relations increase as L2 28 learners spend time studying a language and as their lexicon develops. This 29 finding has important implications for the growth of lexical knowledge and a 30 greater range of lexical meanings. Specifically, it demonstrates that L2 learners 31 develop the ability to use more abstract language as time increases. The find- 32 ings also indicate the growth of L2 lexical networks and increased L2 lexical 33 proficiency. 34 Although the findings of the study are convincing, there are some limi- 35 tations. First and foremost, the number of L2 learners examined was small. 36 Although this is often unavoidable in longitudinal work, future studies should 37 attempt to gather language data from larger groups of L2 learners. Another 38 possible limitation is the lexical size of WordNet. Although WordNet is quite 39 large with over 170,000 lexical items (nouns and verbs only), it is not totally 40 inclusive and some lexical items as well as hypernymic levels might have been 41 omitted. In addition, this study was unable to control for the effects of the

Language Learning 59:2, June 2009, pp. 307–334 328 lang_508 langxml-als.cls (1994/07/13 v1.2u Standard LaTeX document class) 3-4-2009 :626

1 Crossley, Salsbury, and McNamara Hypernymic Relationships 2

3 learners’ L1. This mostly resulted from the methods used in this study, which 4 depended on a large-scale analysis that concentrated not on specific word pro- 5 duction but on the development of the lexical system as a whole. However, as 6 a past study has shown, there is a likely link between the concepts developed 7 in an L1 and their influence on conceptual and lexical development in an L2 8 (Ijaz, 1986). Future studies should consider to what degree an L1 influences 9 the development of sense relations and lexical networks. Additionally, this 10 study was not able to make a strong connection between lexical proficiency 11 and hypernymic development because the study depended on a measure of LD. 12 LD, although an important component of lexical proficiency, is not completely 13 representative of lexical proficiency. Thus, future studies should examine links 14 among lexical proficiency, hypernymy, and lexical diversity. This, however, is 15 beyond the scope of the current study. 16 Despite these limitations, a computational approach such as that found in 17 this study is both necessary and beneficial, as lexical acquisition is a complex 18 phenomenon and the exploration of lexical networks benefits from computa- 19 tional approaches (Meara, 2006). Additionally, the approach found in this article 20 blends real-world data and computational tools, enabling the investigation of 21 lexical development in a way that utilizes natural language processing and relies 22 on both psycholinguistic and corpus linguistic methods. Because the results are 23 encouraging, we contend that future studies in L2 lexical development would 24 benefit from considering hypernymic relations as well as other sense relations 25 and deeper knowledge levels. In addition, more studies are needed that consider 26 natural language data and computational tools. Such approaches allow for the 27 investigation of lexical network formation and provide valuable insights into 28 how L2 learners develop deep lexical knowledge. 29 Revised version accepted 9 April 2008 30 31 32 33 References 34 35 Agirre, E., & Rigaus, G. (1996). Word sense disambiguation using conceptual density. In Proceedings of the 16th international conference on computational linguistics 36 (pp. 16–22). Copenhagen: Cascadilla Press. 37 Anglin, J. M. (1977). Word, object, and conceptual development. New York: Norton. 38 Anglin, J. M. (1993). Vocabulary development: A morphological analysis. 39 Monographs of the Society for Research in Child Development, 58(10), 1–166. Q4 40 Bardovi-Harlig, K. (2000). Tense and aspect in second language acquisition: A study 41 of form, meaning, and use. Oxford: Blackwell.

329 Language Learning 59:2, June 2009, pp. 307–334 lang_508 langxml-als.cls (1994/07/13 v1.2u Standard LaTeX document class) 3-4-2009 :626

1 Crossley, Salsbury, and McNamara Hypernymic Relationships 2

3 Berlin, B., Breedlove, D., & Raven, P. (1974). Principles of Tzeltal plant classification. 4 New York, Academic. 5 Bogaards, P. (2001). Lexical units and the learning of foreign language vocabulary. 6 Studies in Second Language Acquisition, 23, 321–343. 7 Brill, E. (1995). Transformation-based error-driven learning and natural language 8 processing: A case study in part of speech Tagging. Computational Linguistics, 21, 9 543–565. Brown, R. (1958). How shall a thing be called? Psychological Review, 65, Q5 10 Chaffin, R., & Glass, A. (1990). A comparison of hyponym and synonym decisions. 11 Journal of Psycholinguistic Research, 19(4), 265–280. 12 Coltheart, M. (1981). The MRC psycholinguistic database. Quarterly Journal of 13 Experimental Psychology, 33A, 497–505. 14 Crossley, S. A., Louwerse, M. M., McCarthy, P. M., & McNamara, D. S. (2007). A 15 linguistic analysis of simplified and authentic texts. Modern Language Journal, 16 91(2), 15–30. 17 Crossley, S. A., & McNamara, D. S. (in press). Assessing second language reading 18 texts at the intermediate level: An approximate replication of Crossley, Louwerse, 19 McCarthy, and McNamara (2007). Language Teaching. Q6 20 Daller, H., van Hout, R., & Treffers-Daller, J. (2003). Lexical richness in the 21 spontaneous speech of bilinguals. Applied Linguistics, 24(2), 197– 222. 22 Defour, R., & Kroll, J. F. (1995). Matching words to concepts in two languages: A test 23 of the concept mediation model of bilingual representation. Memory & Cognition, 24 23, 166–180. 25 de la Fuente, M. J. (2002). Negotiation and oral acquisition of L2 vocabulary: The 26 roles of input and output in the receptive and productive acquisition of words. 27 Studies in Second Language Acquisition, 24, 81–112. 28 Downing, P. (1980). Factors influencing lexical choice in narrative. In W. Chafe (Ed.), 29 The pear stories: Cognitive, cultural, and linguistic aspects of narrative production 30 (pp. 89–126). Norwood, NJ: Ablex 31 Ellis, N. C. (in press). Optimizing the input: Frequency and sampling in usage-based 32 and form-focused learning. In M. H. Long & C. Doughty (Eds.), Handbook of 33 second and foreign language teaching. Oxford: Blackwell. Q7 Ellis, N. C., & Larsen-Freeman, D. (2006). Language Emergence: Implications for 34 Applied Lingustics. Applied Linguistics, 27(4), 558–589. 35 Ellis, R. (1994). A theory of instructed second language acquisition. In N. C. Ellis 36 (Ed.), Implicit and Explicit Learning of Languages (pp. 79–114). London: 37 Academic Press. 38 Ellis, R. (1995). Modified oral input and the acquisition of word meanings. Applied 39 Linguistics, 16, 409–435. 40 Ellis, R., Tanaka, Y., & Yamakazi, A. (1994). Classroom interaction, comprehension, 41 and L2 vocabulary acquisition. Language Learning, 44, 449–491.

Language Learning 59:2, June 2009, pp. 307–334 330 lang_508 langxml-als.cls (1994/07/13 v1.2u Standard LaTeX document class) 3-4-2009 :626

1 Crossley, Salsbury, and McNamara Hypernymic Relationships 2

3 Fellbaum, C. (1998). WordNet: An electronic lexical database. Cambridge, MA: MIT 4 Press. 5 Ferrer i Cancho, R., & Sole,´ R. V.(2001). The small world of human language. 6 Proceedings of the Royal Society of London B, 268, 2261–2265. 7 Ferrer i Cancho, R., Sole,´ R. V., & Kohler,¨ R. (2004). Patterns in syntactic dependency 8 networks. Physical Review, E69, 0519151–0519158. 9 Fragos, K., & Maistros, Y. (2005). Distributional analysis of related synsets in WordNet for a word sense disambiguation task. International Journal on Artificial 10 Intelligence Tool, 14(6), 919–034. Q8 11 Gilhooly, K. J., & Logie, R. H. (1980). Age-of-acquisition, imagery, concreteness, 12 familiarity, and ambiguity measures for 1,944 words. Behavior Research Methods 13 & Instrumentation, 12(4), 395–427. 14 Graesser, A. C., McNamara, D. S., Louwerse, M. M., & Cai, Z. (2004). Coh-Metrix: 15 Analysis of text on cohesion and language. Behavioral Research Methods, 16 Instruments, and Computers, 36, 193–202. 17 Haastrup, K., & Henriksen, B. (2000). Vocabulary acquisition: Acquiring depth of 18 knowledge through network building. International Journal of Applied Linguistics, 19 10(2), 221–240. 20 Herrmann, D. J., & Chaffin, R. (1986). Comprehension of semantic relations as a 21 function of definitions of relations. In F. Klix (Ed.), Learning and memory (pp. 311–319). Berlin: North-Holland. 22 Huckin, T., & Coady, J. (1999). Incidental vocabulary acquisition in a second 23 Language. Studies in Second Language Acquisition, 21(2), 181–193. 24 Ijaz, I. H. (1986). Linguistic and cognitive determinants of lexical acquisition in a 25 second language. Language Learning, 36(4), 401–451. 26 Jarvis, S. (2002). Short texts, best fitting curves and new measures of lexical diversity. 27 Language Testing, 19(1), 57–84. 28 Jurafsky, D., & Martin, J. H. (2000). Speech and language processing: An introduction 29 to natural language processing, computational linguistics, and speech recognition, 30 Upper Saddle River, NJ: Prentice-Hall. 31 Klix, F., & van der Meer, E. (1980). The method of analogy recognition of the 32 determination of semantic relations in long-term memory. In F. Klix & J. Hoffman 33 (Eds.), Cognition and memory (pp. 145–152). Berlin: North-Holland. Lakoff, G. (1987). Women, fire, and dangerous things. Chicago: University of Chicago 34 Press. 35 Leacock, C., & Chodorow, M. (1998). Combining local context and WordNet 36 similarity for word sense disambiguation. In C. Fellbaum (Ed.), WordNet: An 37 electronic lexical database (pp. 265–283). Cambridge: MIT Press. 38 Levenston, E., & Blum, S. (1977). Aspects of lexical simplification in the speech and 39 writing of advanced adult learners. In P. S. Corder & E. Roulet (Eds.), The notions 40 of simplification, interlanguages and pidgins and their relation to second language 41 pedagogy (pp. 51–72). Librairie Droz. Q9

331 Language Learning 59:2, June 2009, pp. 307–334 lang_508 langxml-als.cls (1994/07/13 v1.2u Standard LaTeX document class) 3-4-2009 :626

1 Crossley, Salsbury, and McNamara Hypernymic Relationships 2

3 LeVine, R. A. (1980). Influences of women’s schooling on maternal behavior in the 4 third world. Comparative Education Review, 24, 78–105. 5 Malvern, D., & Richards, B. (2002). Investigating accommodation in language 6 proficiency interviews using a new measure of lexical diversity. Language Testing, 7 19(1), 85–104. 8 Malvern, D. D., Richards, B. J., Chipere, N., & Duran, P. (2004). Lexical diversity and 9 language development: Quantification and assessment. Houndmills, UK: Palgrave Macmillan. 10 McCarthy, P. M. (2005). An assessment of the range and usefulness of lexical diversity 11 measures and the potential of the measure of textual, lexical diversity (MTLD). 12 Dissertation Abstracts International, 66(12). (UMI No. 3199485) 13 Meara, P. (1982). Word association in a foreign language: A report on the Birkbeck 14 Vocabulary Project. Nottingham Linguistic Circular, 11, 20–37. 15 Meara, P. (2002). The rediscovery of vocabulary. Second Language Research, 18(4), 16 393–407. 17 Meara, P. (2006). Emergent properties of multilingual lexicons. Applied Linguistics, 18 27(4), 620–644. 19 Mervis, C. B., & Crisafi, M. A. (1982). Order of acquisition of subordinate-, basic-, 20 and superordinate-level categories. Child Development, 53, 258–266. 21 Miller, G. A., Beckwith, R., Fellbaum, C., Gross, D., & Miller, K. (1990). Five papers on WordNet. Cognitive Science Laboratory, Princeton University, No. 43. Q10 22 Miller, G. A., & Teibel, D. A. (1991). A proposal for lexical disambiguation. In (Eds.), 23 Human language technology conference: Proceedings of the workshop on speech 24 and natural language (pp. 395–399). Pacific Grove, CA: Association for 25 Computational Linguistics. Q11 26 Miller, J. (1981). Assessing language production in children: Experimental 27 procedures. Austin, TX: Allyn & Bacon. 28 Murphy, G. L. (2004). The big book of concepts. Cambridge, MA: MIT Press. 29 Nation, I. S. P. (2005). Teaching and learning vocabulary. In E. Hinkel (Ed.), 30 Handbook of research in second language teaching and learning (pp. 581–595). 31 Mahwah, NJ: Lawrence Erlbaum. 32 Ordonez, C. L., Carlo, M. S., Snow, C. E., & McLaughlin, B. (2002). Depth and 33 breadth of vocabulary in two languages: Which vocabulary skills transfer. Journal of Educational Psychology, 94(4), 719–728. 34 Paivio, A., Yuille, J. C., & Madigan, S. (1968). Concreteness, imagery, and 35 meaningfulness values for 925 nouns. Journal of Experimental Psychology 36 Monograph Supplement, 76, 1–25. 37 Piaget, J. (1972). The principles of genetic epistemology. New York: Basic Books. 38 Polio, C. (2001). Research methodology in second language writing research: The case 39 text-based studies. In T. Silva & P. K. Matsuda (Eds.), On second language writing 40 (pp. 91–115). Mahwah, NJ: Lawrence Erlbaum. 41

Language Learning 59:2, June 2009, pp. 307–334 332 lang_508 langxml-als.cls (1994/07/13 v1.2u Standard LaTeX document class) 3-4-2009 :626

1 Crossley, Salsbury, and McNamara Hypernymic Relationships 2

3 Resnik, P. (1992). WordNet and distributional analysis: A class-based approach to 4 lexical discovery statistically-based natural-language-processing techniques. In 5 Proceedings of the AAAI Workshop on Statistically-Based NLP 6 Techniques. Q12 7 Rintell, E. M. (1989). That reminds me of a story: The use of language to express 8 emotion by second-language learners and native speakers. In M. R. Eisenstein 9 (Ed.), The dynamic interlanguage (pp. 237–257). New York: Plenum Press. 10 Rosch, E. (1973). On the internal structure of perceptual and semantic categories. 11 In T. E. Moore (Ed.), Cognitive development and the acquisition of language 12 (pp. 111–144). New York: Academic Press. 13 Rosch, E., Mervis, C., Gray, W., Johnson, D., and Boyes-Braem, P. (1976). Basic 14 objects in natural categories. Cognitive Psychology, 8, 573–605. 15 Sharifian, F. (2002). Processing hyponymy in L1 and L2. Journal of Psycholinguistic 16 Research, 31(4), 421–436. 17 Snow, C. (1988). The problem with bilingual education research critiques: A response 18 to Rossell. Equity and Excellence, 23(4), 30–31. 19 Snow, C. E. (1990). The development of definitional skill. Journal of Child Language, 20 17, 697–710. 21 Snow, C. E., Cancino, H., De Temple, J., & Schley, S. (1991). Giving formal definitions: Linguistic or metalinguistic skill? In E. Bialystok (Ed.), Language 22 processing in bilingual children (pp. 90–112). Cambridge: Cambridge University 23 Press. 24 Snow, C. E., Cancino, H., Gonzalez, P., & Shriberg, E. (1989). Giving formal 25 definitions: An oral language correlate of school literacy. In D. Bloome (Ed.), 26 Classrooms and literacy (pp. 233–249). Norwood, NJ: Ablex. 27 Soderman,¨ T. (1993). Word associations of foreign language learners and native 28 speakers: The phenomenon of a shift in response type and its relevance for lexical 29 development. In H. Ringbom (Ed.), Near-native proficiency in English (pp. 30 91–182). Abo, Finland: Abo Akademi. 31 Templin, M. (1957). Certain language skills in children. Minneapolis: University of 32 Minnesota Press. 33 Toglia, M. P., & Battig, W. F. (1978). Handbook of semantic word norms. Hillsdale, NJ: Lawrence Erlbaum. 34 Tversky, B., & Hemenway, K. (1984). Objects, parts, and categories. Journal of 35 Experimental Psychology: General, 113, 169–193. 36 Verhallen, M., & Schoonen, R. (1993). Lexical knowledge of monolingual and 37 bilingual children. Applied Linguistics, 14, 344–363. 38 Verhallen, M., & Schoonen, R. (1998). Lexical knowledge in L1 and L2 third and fifth 39 graders. Applied Linguistics, 19, 452–470. 40 Vygotsky, L. (1962). Thought and language. Cambridge, MA: MIT Press. 41

333 Language Learning 59:2, June 2009, pp. 307–334 lang_508 langxml-als.cls (1994/07/13 v1.2u Standard LaTeX document class) 3-4-2009 :626

1 Crossley, Salsbury, and McNamara Hypernymic Relationships 2

3 Welch, L., & Long, L. (1940). The higher structural phases of concept formation of 4 children. Journal of Psychology, 9, 59–95. 5 Wesche, M., & Paribakht, T. M. (1996). Assessing vocabulary knowledge: Depth vs. 6 breadth. Canadian Modern Language Review, 53, 13–40. 7 Wolter, B. (2001). Comparing the L1 and L2 mental lexicons: A depth of individual 8 word knowledge model. Studies in Second Language Acquisition, 23, 41–70. 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41

Language Learning 59:2, June 2009, pp. 307–334 334 lang_508 langxml-als.cls (1994/07/13 v1.2u Standard LaTeX document class) 3-4-2009 :626

Queries

Q1 Author: Please update N. Ellis, in press, throughout text and in the References. Q2 Author: Please update Crossley & McNamara, in press, throughout text and in the References. Q3 Author: Should TTR be spelled out at ﬁrst use? Q4 Author: Book? Journal? Please clarify. Q5 Author: Please provide the page range for the chapter. Q6 Author: Page range needed. Q7 Author: Please provide the page range for the chapter. Q8 Author: Please check the page range. 919–934 or 919–1034? Q9 Author: Please provide the location of the publisher. Q10 Author: Report? Monograph? Please clarify. Q11 Author: Please provide the editors. Q12 Author: Please provide the month and location of the Workshop.