A model of rapid phonotactic generalization

Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 1126–1131, Lisbon, Portugal, 17-21 September 2015. c⃝2015 Association for Computational Linguistics. A model of rapid phonotactic generalization Tal Linzen Department of Linguistics New York University linzen@nyu.edu Timothy J. O’Donnell Brain and Cognitive Sciences Massachusetts Institute of Technology timod@mit.edu Abstract The phonotactics of a language describes the ways in which the sounds of the language combine to form possible mor- phemes and words. Humans can learn phonotactic patterns at the level of abstract classes, generalizing across sounds (e.g., “words can end in a voiced stop”). More- over, they rapidly acquire these general- izations, even before they acquire sound- specific patterns. We present a probabilis- tic model intended to capture this early- abstraction phenomenon. The model rep- resents both abstract and concrete gen- eralizations in its hypothesis space from the outset of learning. This—combined with a parsimony bias in favor of compact descriptions of the input data—leads the model to favor rapid abstraction in a way similar to human learners. 1 Introduction Natural languages place restrictions on the ways in which sounds can combine to form words (the phonotactics of the language). The velar nasal [N], for example, occurs at the end of English sylla- bles, as in ring [ôIN] or finger [fINg@ô], but never at the beginning of a syllable: English does not have words like *ngir [NIô]. English speakers are aware of this constraint, and judge forms that start with a [N] as impossible English words. Sounds that share articulatory and/or perceptual properties often have similar phonotactic distribu- tions. German, for example, allows voiced obstru- ents,1 such as [b] and [g], to occur anywhere in the word except at its end: [bal] is a valid German word, but [lab] isn’t. Speakers use such features of sounds to form phonotactic generalizations, which can then apply to sounds that do not appear in their language. Al- though no English words start with either [sô] or [mb], English speakers judge srip to be a better po- tential word of English than mbip (Scholes, 1966); this is likely because [sô] shares properties with strident-liquid clusters that do exist in English, such as [sl] as in slip and [Sô] as in shrewd, whereas [mb] does not benefit from any sonorant-stop on- set sequences (*[nt])—none exist in English. Recent studies have investigated how humans acquire generalizations over phonological classes in an artificial language paradigm (Linzen and Gallagher, 2014; Linzen and Gallagher, 2015). The central finding of these studies was that participants rapidly learned abstract phonotactic constraints and exhibited evidence of generaliza- tions over classes of sounds before evidence of phoneme-specific knowledge. This paper presents a probabilistic genera- tive model, the Phonotactics As Infinite Mixture (PAIM) model, which exhibits similar behavior. This behavior arises from the combination of two factors: the early availability of abstract phono- logical classes in the learner’s hypothesis space; and a parsimony bias implemented as a Dirichlet process mixture, which favors descriptions of the data using a single pattern over …

Content truncated to 500 words