Proceedings of the Society for Computation in Linguistics Volume 4 Article 19 2021 Consonant harmony, disharmony, memory and time scales Adamantios Gafos Universität Potsdam, [email protected] Follow this and additional works at: https://scholarworks.umass.edu/scil Part of the Cognition and Perception Commons, Computational Linguistics Commons, Phonetics and Phonology Commons, and the Typological Linguistics and Linguistic Diversity Commons Recommended Citation Gafos, Adamantios (2021) "Consonant harmony, disharmony, memory and time scales," Proceedings of the Society for Computation in Linguistics: Vol. 4 , Article 19. DOI: https://doi.org/10.7275/vhf9-qh54 Available at: https://scholarworks.umass.edu/scil/vol4/iss1/19 This Paper is brought to you for free and open access by ScholarWorks@UMass Amherst. It has been accepted for inclusion in Proceedings of the Society for Computation in Linguistics by an authorized editor of ScholarWorks@UMass Amherst. For more information, please contact [email protected]. Consonant harmony, disharmony, memory and time scales Adamantios Gafos University of Potsdam [email protected] Abstract The four-way contrast in the distribution of ejectivization in a hypothetical input set {/k'ap'i/, I argue that properties of memory, a so far /k'api/, /kap'i/, /kapi/} is neutralized to a three- largely neglected source of explanation in way contrast. This is because the high ranking phonological patterns, offer a grounding for a number of not well understood traits *1vs2 constraint bans candidate sets exhibiting a of long distance consonantal restrictions contrast between one versus two ejectives as in set (both dissimilatory and assimilatory). (1i), and neutralization to one ejective as in (1ii) is preferred over neutralization to two ejectives as 1 Introduction in (1iii), because the faithfulness constraint to the ejectivization feature FAITH(+CG) outranks the In the great variety of long distance consonantal *0vs1 constraint which bans candidate sets restrictions met across languages, production and exhibiting a contrast between zero versus one perception factors have been implicated. Here, I consider the neglected factor of memory. I argue ejective (‘+CG’ stands for constricted glottis). that properties of memory, when combined with the other two factors, enable new answers to {/k'ap'i/, /k'api/, unresolved questions about long distance /kap'i/, /kapi/} *1vs2 FAITH(+CG) *0vs1 consonantal restrictions. Section 2 considers short i. term memory in dissimilatory restrictions. Long [k'ap'i, k'api, term memory and assimilatory restrictions are kap'i, kapi] ** **** taken up in Section 3. Section 4 concludes. ii. 2 Dissimilation [k'api, kap'i, * ** Dissimilations involving laryngeal features are a kapi] primary example of long distance co-occurrence iii. restrictions in consonants (MacEachern, 1999; [k'ap'i, kapi] Mackenzie, 2009; Gallagher, 2010). A number of **! languages ban the presence of repeated ejectives, aspirated or implosive consonants within roots. Table 1. *1vs2 >> FAITH(+CG) >> *0vs1 For example, in Shuswap, roots cannot contain two ejectives, e.g., using C' to denote In support of the ranking *1vs2 >> *0vs1 that ejectivization, /kʷ'alt/ ‘to stagger’, /qet'/ ‘to hoist’, plays out in this loss of contrast analysis, /kʷup/ ‘to push’, /qmut/ ‘hat’, but */kʷ'alt'/, */q'et'/ Gallagher conducted discrimination experiments and so on are unattested (Kuipers, 1974). where listeners heard pairs of CVCV stimuli and Gallagher (2010) proposes a contrast gave a same versus different judgment. The neutralization analysis of such restrictions. The stimuli were disyllabic ‘words’ containing zero, basic generalization captured in that analysis is one, or two ejectives and were paired in three that roots with two ejectives, as in */k'ap'i/, are conditions: condition 1vs2, that is, one versus two unattested because they may not contrast with instances of the ejective as in [kap'i]-[k'ap'i], 0vs1 roots with only one ejective, as in /k'api/ or /kap'i/. as in [kapi]-[k'api], and 0vs2 as in [kapi]-[k'ap'i] An example from this analysis is in tableau (1). (and so on for aspirated Cs). Percent of judgment 188 Proceedings of the Society for Computation in Linguistics (SCiL) 2021, pages 188-205. Held on-line February 14-19, 2021 correct (of the same versus different judgment) the number of features. In our case, N is 1 as all was the dependent variable. The main result was pairs in this task (e.g., [kapi-k'api]) are identical that, for both ejectivization and aspiration, the except for [+CG] (thus, the weight parameter 푏푘 1vs2 condition was ‘more difficult’ (i.e., correct can also be set to 1 since effectively there is no responses to same versus different were less other feature, given the stimuli in this task). This accurate) than the 0vs1 condition, which in turn means that the distance between the two stimuli was more difficult than the 0vs2 condition. in any pair of stimuli of the task is given simply Gallagher (2010), whose primary concern is by the number of mismatches with respect to the synchronic analyses of dissimilation patterns as [+CG] feature. Thus, 푑([kapi], [k'api]) = 1 and they play out in their language-particular details, 푠([kapi], [k'api]) = 푒−푑 = 푒−1 = 0.37. The same leaves open the source of these discrimination holds for the other pairs, [kapi-kap'i], [k'api-kapi], asymmetries (‘A major question … is the source [kap'i-kapi]. For each pair, the distance is 1 and of the perceptual asymmetries’, ibid.: 106). therefore the similarity is 푒−1 = 0.37. Thus, the Consider what the task in the experiments average similarity between the two stimuli across above involves. A chain of auditory processing, all four stimuli pairs in this 0vs1 condition is 0.37. encoding of the stimuli in short term (or working) In the 1vs2 condition, the stimuli are [k'ap'i- memory, along with a discrimination decision is k'api], [kap'i-k'ap'i], [k'ap'i-kap'i], [k'api-k'ap'i]. implicated. That is, comparison of the stimuli is These become after overwrite [k'ap'i-k'api] → mediated by representations of their forms in [kap'i-k'api], with a resulting similarity s([kap'i], short term memory (Baddeley, 1986). This invites [k'api]) = 푒−2 = 0.13 (푑 = 2, s = 푒−푑 = 0.13), consideration of the extent to which the results [kap'i-k'ap'i] → [kap'i-kap'i] (푑 = 0, s = 푒−푑 = may derive from properties of short term memory. 1), [k'ap'i-kap'i] → [kap'i-kap'i] (푑 = 0, s = 푒−푑 A well-replicated finding about short term = 1), and finally [k'api-k'ap'i] → [k'api-kap'i] memory is that features of items in a list that (푑 = 2, s = 푒−푑 = 0.13). Across all four stimuli match features of other items tend to be omitted, pairs, the average similarity between the two a so-called interference effect. The mechanism 1 stimuli within pairs is [0.13 + 1 + 1 + 0.13] = which gives rise to this effect is dubbed feature 4 overwrite in models of short term memory 0.57 (higher than for the 0vs1 condition). (Nairne, 1990; Nairne, 2001): in a list of elements These illustrations assume that interference (e.g., syllables), a feature F of the memory trace applies in every experimental trial. To prove a of an element is omitted (with some probability; general result, I assume that overwrite applies more on this below) if it matches the feature of a with probability 푝 and show how to derive the following, adjacent element (Nairne, 1990: 252). discriminability asymmetries in the above Consider how interference plays out in the experiments for any 푝. I begin by expressing the different conditions of Gallagher’s experiments. expected similarity for each of the above pairs as In the 0vs1 condition, the stimuli pairs are [kapi- a function of this probability 푝 of overwrite (1 − k'api], [kapi-kap'i], [k'api-kapi], and [kap'i-kapi]. 푝, no overwrite). In the 1vs2 condition, the No pair is subject to feature overwrite, as the expected distance for [k'ap'i-k'api] → [kap'i-k'api] ejective is not repeated. Discriminability of the is 1(1 − 푝) + 2푝 = 1 + 푝 (read: 1, the distance two words in each pair is function of their when overwrite does not apply, multiplied by the similarity. The similarity 푠(푖, 푗) between words probability of no overwrite (1 − 푝), plus 2, the 푖, 푗 is related to their featural distance 푑(푖, 푗) by distance when overwrite applies, multiplied by 푠(푖, 푗) = 푒−푑(푖,푗) (Shepard, 1987); the larger the the probability of overwrite 푝), 1(1 − 푝) + 0푝 = distance, the less the similarity. Distance is a 1 − 푝 for [kap'i-k'ap'i] → [kap'i-kap'i], 1(1 − function of the number of featural mismatches 푝) + 0푝 = 1 − 푝 for [k'ap'i-kap'i] → [kap'i- between the two forms. More precisely, distance kap'i], and 1(1 − 푝) + 2푝 = 1 + 푝 for [k'api- ∑ 푏푘푀푘 k'ap'i] → [k'api-kap'i]. Hence, the average is given by 푑(푖, 푗) = where 푏푘 is the 푁 1 −푝−1 similarity across the four pairs is [2푒 + weight of the particular feature (some features 4 1 may contribute more to distance than others), 푀 2푒푝−1] = 푒−1 ( 푒−푝 + 푒푝). This expression is 푘 2 is a counter of the featural mismatches, and N is lower bounded (at 푝 = 0) by the similarity of the 189 −1 F 0vs1 condition, 푒 = 0.37 which does not vary swap positions: /…Ci …Cj…/ (read: feature F is as a function of 푝. Figure 1 plots the similarities associated to consonant in position j) changes to F across these conditions.