Diana Apoussidou
Modeling the acquisition of French liaison using allomorphy
As language acquisition research shows (e.g. Chevrot et al. 2008), children learning French are creative when segmenting nouns starting with a vowel. Words like arbre ‘tree’ are in adult speech rarely produced in isolation and undergo a liaison with the final consonant of the preceding word, e.g. un arbre is pronounced as oe.narbr, or des arbre as de.zarbr. Children until the age of 4;6 therefore produce errors such as narbr or zarbr. Chevrot et al. (2008) analyze these errors in terms of templates that the children use in the course of development. The templates are made up of un+/Nword2/ or deux+/Zword2/ etc., where the extra consonant in front of a word depends on the preceding word. I propose instead that the errors produced by the children can be analyzed in terms of allomorphy: children hypothesize different underlying representations for words (e.g. literally /arbr/, /narbr/ and /zarbr/ for ‘tree’) depending on what they can observe. This can be modeled with an optimization-based grammar where different underlying forms of a word are represented by lexical constraints. The results show that even with a resulting ‘correct’ lexicon (e.g. vowel-initial /arbr/ as underlying representation of ‘tree’), interference with the grammar can lead to the use of allomorphs in production (e.g. /narbr/ in combination with un, yielding /oe#narbr/ instead of /oen#arbre).
Shigeto Kawahara, Rutgers University
Probing knowledge of similarity through puns
*please note- this talk will be held from 4:00-6:00pm in 32-D831*
This talk outlines the aims, results and future prospects of a general research program which investigates knowledge of similarity through the investigation of Japanese imperfect puns, dajare. I argue that speakers attempt to maximize the similarity between corresponding segments in composing puns, just as in phonology where speakers maximize the similarity between, for example, inputs and outputs. In this sense, we find non-trivial parallels between phonology and pun patterns. I further argue that we can take advantage of these parallels, and use puns to investigate our linguistic knowledge of similarity. To develop these arguments, I start with an overview of the results of some recent projects, and follow that with patterns that provide interesting lines of future research.
MUMM Practice talks
Hyesun Cho
The problem of generalization in a statistical learning model of phonotactics
Achieving a proper level of generality in OT-style constraints has not been a problem in most of the phonological grammar learning models up to now (such as GLA (Boersma, 1997); a Maximum Entropy model (Goldwater and Johnson, 2003); a GLA-HG model (Coetzee and Pater, 2008)), because in those models, constraints do not have to be learned. Instead, learning a grammar mainly involves finding rankings or weights of the given set of constraints. In contrast, the phonotactic learning model of Hayes and Wilson (2008) is different from others in that the model derives constraints themselves from the surface forms in the training data. This paper shows that in doing so, it is not trivial to learn constraints with a proper level of generalization.
I ran the Hayes and Wilson model on Korean training data. In the resulting grammar, the model makes overly-broad generalizations, i.e., the constraints that penalize both possible (grammatical) and impossible (ungrammatical) sequences equally, especially when the frequency differences between the possible sequences and the impossible sequences are very small. Because of this, some of the possible sequences are predicted to be worse than impossible sequences. The three problem cases include: post-obstruent tensing, diphthong restrictions, and labials-[?] sequences. I discuss two possible solutions: adjusting feature specifications and employing an additional learning bias.
The grammar learned by the Hayes and Wilson model consists of markedness constraints only. I ran a model that uses both faithfulness constraints and markedness constraints (Goldwater and Johnson, 2003) for the problem cases above. It turns out that a proper level of generalization in the constraints is crucial in a faithfulness model as well.
Patrick Jones
Evidence for the Phonological Stem in Kinande
A number of recent studies on Bantu verbal phonology (Downing 1994, 1998, 1999, 2000; Herman 1996) have argued that the phonological processes which affect the Bantu verb, rather than referring directly to constituents of morpho-syntactic structure (M-Constituents), instead refer exclusively to morpho-prosodic constituents (P-Constituents) which are derived from them. Evidence for this view, which has its origins in the work of Selkirk (1986) and Inkelas (1989), is found in a number of phonological processes whose domains of application approximate constituents defined by morpho-syntactic structure, but do not match them exactly. In this talk, I will argue that the existence of a Phonological Stem (PStem) domain is strongly supported by the verbal phonology of Kinande, and that by recognizing the PStem as a morpho-prosodic domain to which phonological constraints can refer, it is possible to provide straightforward and unified analyses of four distinct phonological processes - Verbal Reduplication, Intonational Tone Assignment, Lexical Tone Assignment, and Purposive Suffix Affixation - that must otherwise be explained in terms of arbitrary and idiosyncratic constraints. However, I will argue against the position that only morpho-prosodic constituents, and not morpho-syntactic ones, may be referred to by phonological processes. I will argue that reference to the MStem as well as reference to the PStem is necessary in order to successfully account for Verbal Reduplication and Lexical Tone Assignment, and will therefore argue for a theory in which phonological processes make necessary, but not exclusive, reference to morpho-prosodic domains.
