Measuring the perceptual availability of phonological features during language acquisition using unsupervised binary stochastic autoencoders

Proceedings of NAACL-HLT 2019, pages 69–85 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 69 Measuring the Perceptual Availability of Phonological Features During Language Acquisition Using Unsupervised Binary Stochastic Autoencoders Cory Shain Department of Linguistics The Ohio State University shain.3@osu.edu Micha Elsner Department of Linguistics The Ohio State University melsner@ling.osu.edu Abstract In this paper, we deploy binary stochastic neu- ral autoencoder networks as models of infant language learning in two typologically unre- lated languages (Xitsonga and English). We show that the drive to model auditory per- cepts leads to latent clusters that partially align with theory-driven phonemic categories. We further evaluate the degree to which theory- driven phonological features are encoded in the latent bit patterns, finding that some (e.g. [±approximant]), are well represented by the network in both languages, while others (e.g. [±spread glottis]) are less so. Together, these findings suggest that many reliable cues to phonemic structure are immediately available to infants from bottom-up perceptual charac- teristics alone, but that these cues must eventu- ally be supplemented by top-down lexical and phonotactic information to achieve adult-like phone discrimination. Our results also suggest differences in degree of perceptual availabil- ity between features, yielding testable predic- tions as to which features might depend more or less heavily on top-down cues during child language acquisition. 1 Introduction Distinctive features like [±voice] and [±sonorant] have been a core construct of phonological the- ory for many decades (Trubetskoy, 1939; Jakob- son et al., 1951; Chomsky and Halle, 1968; Clements, 1985). They have been used in au- tomatic speech recognition (Livescu and Glass, 2004), and psycholinguistic evidence suggests that they are cognitively available during language ac- quisition (Kuhl, 1980; White and Morgan, 2008). Nonetheless, distinctive features are not directly observed by humans; they are abstractions that must be inferred from dense perceptual infor- mation (sound waves) during language acquisi- tion and comprehension, which raises questions about how they are learned and recognized. In adults, phonological comprehension is aided by top-down lexical and phonotactic (i.e. sound se- quencing) constraints. For example, the classic phonemic restoration effect (Warren, 1970) shows that adults infer missing phonemes from context with such ease that they often fail to notice when acoustic cues to phone identity are erased. How- ever, infants first learning their phonemic cate- gories have not yet acquired reliable top-down lex- ical and phonotactic models and must rely more heavily on bottom-up perceptual information. To a learner faced with the immense challenge of dis- covering structure in dense perceptual input, do theory-driven phonological features “stand out” or are they swamped by noise? In this paper, we ad- dress this question using an unsupervised compu- tational acquisition model. Previous models of phonological category in- duction have emphasized the importance of top- down information (information about the con- texts in which phonemes occur) (Peperkamp et al., 2006; Swingley, 2009; Feldman et al., 2009a, 2013a,b; Moreton and Pater, 2012a,b; Martin et al., 2013; Pater and Moreton, 2014; Frank et al., 2014; Doyle et al., 2014; …

Content truncated to 500 words