Sign constraints on feature weights improve a joint model of word segmentation and phonology

Human Language Technologies: The 2015 Annual Conference of the North American Chapter of the ACL, pages 303–313, Denver, Colorado, May 31 – June 5, 2015. c⃝2015 Association for Computational Linguistics Sign constraints on feature weights improve a joint model of word segmentation and phonology Mark Johnson Macquarie University Sydney, Australia Mark Johnson@MQ.edu.au Joe Pater University of Massachusetts, Amherst Amherst, MA, USA pater@linguist.umass.edu Robert Staubs University of Massachusetts, Amherst Amherst, MA, USA rstaubs@linguist.umass.edu Emmanuel Dupoux ´Ecole des Hautes Etudes en Sciences Sociales, ENS, CNRS, Paris, France emmanuel.dupoux@gmail.com Abstract This paper describes a joint model of word segmentation and phonological alternations, which takes unsegmented utterances as input and infers word segmentations and underlying phonological representations. The model is a Maximum Entropy or log-linear model, which can express a probabilistic version of Opti- mality Theory (OT; Prince and Smolensky (2004)), a standard phonological framework. The features in our model are inspired by OT’s Markedness and Faithfulness constraints. Fol- lowing the OT principle that such features in- dicate “violations”, we require their weights to be non-positive. We apply our model to a modified version of the Buckeye corpus (Pitt et al., 2007) in which the only phonological alternations are deletions of word-final /d/ and /t/ segments. The model sets a new state-of- the-art for this corpus for word segmentation, identification of underlying forms, and identi- fication of /d/ and /t/ deletions. We also show that the OT-inspired sign constraints on fea- ture weights are crucial for accurate identifi- cation of deleted /d/s; without them our model posits approximately 10 times more deleted underlying /d/s than appear in the manually annotated data. 1 Introduction This paper unifies two different strands of research on word segmentation and phonological rule induc- tion. The word segmentation task is the task of segmenting utterances represented as sequences of phones into sequences of words. This is an ideal- isation of the lexicon induction problem, since the resulting words are phonological forms for lexical entries. In its simplest form, the data for a word segmen- tation task is obtained by looking up the words of an orthographic transcript (of, say, child-directed speech) in a pronouncing dictionary and concate- nating the results. However, this formulation sig- nificantly oversimplifies the problem because it as- sumes that each token of a word type is pronounced identically in the form specified by the pronouncing dictionary (usually its citation form). In reality there is usually a significant amount of pronunciation vari- ation from token to token. The Buckeye corpus, on which we base our exper- iments here, contains manually-annotated surface phonetic representations of each word as well as the corresponding underlying form (Pitt et al., 2007). For example, a token of the word “lived” has the underlying form /l.ih.v.d/ and could have the sur- face form [l.ah.v] (we follow standard phonological convention by writing underlying forms with slashes and surface forms with square brackets, and use the Buckeye transcription format). There is a large body of work in the phonolog- ical literature on inferring phonological rules …

Content truncated to 500 words