Using LSTMs to Assess the Obligatoriness of Phonological Distinctive Features for Phonotactic Learning

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1595–1605 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 1595 Using LSTMs to Assess the Obligatoriness of Phonological Distinctive Features for Phonotactic Learning Nicole Mirea∗ and Klinton Bicknell‡, ∗ ∗Northwestern University ‡Duolingo nimirea@u.northwestern.edu klinton@duolingo.com Abstract To ascertain the importance of phonetic in- formation in the form of phonological dis- tinctive features for the purpose of segment- level phonotactic acquisition, we compare the performance of two recurrent neural network models of phonotactic learning: one that has access to distinctive features at the start of the learning process, and one that does not. Though the predictions of both models are significantly correlated with human judgments of non-words, the feature-naive model signif- icantly outperforms the feature-aware one in terms of probability assigned to a held-out test set of English words, suggesting that distinc- tive features are not obligatory for learning phonotactic patterns at the segment level. 1 Introduction Knowing a language involves having systematic expectations about the sequential sound patterns within syllables and words in the language— a sensitivity to the phonotactic generalizations that exist in the language. This sensitivity helps language users segment a continuous stream of speech (Vitevitch et al., 1997), incorporate new words into the lexicon (Storkel et al., 2006), and reconstruct parts of an utterance that may have been obscured by noise. However, the details of how language learners infer these phonotactic gen- eralizations from incoming acoustic data are still unclear. The current project seeks to clarify the ex- tent to which phonetic information (at the level of phonological distinctive features) is useful for pre- dicting upcoming phones within a word, by build- ing computational models of phonotactic acquisi- tion. Phonotactic patterns are typically stated in terms of generalizations over natural classes; for example, voiced stops cannot follow voiceless stops word-finally in English. These natural classes are defined by a hierarchy or set of dis- tinctive features that is either taken to be univer- sal across languages (Chomsky and Halle, 1965; Clements, 2009) or emergent from the process of phonological acquisition—including phonotac- tic acquisition—in a particular language (Mielke, 2008; Dresher, 2015). Nevertheless, most models of phonotactic acquisition require that phonologi- cal distinctive features be specified in advance of learning. Our work interrogates this assumption through the following questions: 1. Is external information regarding phonologi- cal distinctive features a necessary prerequi- site for learning word-level phonotactic gen- eralizations? 2. Must models become sensitive to phonologi- cal properties of incoming segments in order to represent phonotactic generalizations? To answer them, we use recurrent neural net- works with long short-term memory (LSTM) nodes, which have shown considerable success in learning patterns at the word (Sundermeyer et al., 2015) and character (Kim et al., 2016) levels. These models encode each phonetic segment in the inventory as a vector of numbers. With ex- posure to more training data, these representations adapt to the task at hand: incrementally predicting each segment in a word, given all previous …

Content truncated to 500 words