A Joint Learning Model of Word Segmentation, Lexical Acquisition, and Phonetic Variability

Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 42–54, Seattle, Washington, USA, 18-21 October 2013. c⃝2013 Association for Computational Linguistics A Joint Learning Model of Word Segmentation, Lexical Acquisition, and Phonetic Variability Micha Elsner melsner0@gmail.com Dept. of Linguistics The Ohio State University Sharon Goldwater sgwater@inf.ed.ac.uk ILCC, School of Informatics University of Edinburgh Naomi H. Feldman nhf@umd.edu Dept. of Linguistics University of Maryland Frank Wood fwood@robots.ox.ac.uk Dept. of Engineering University of Oxford Abstract We present a cognitive model of early lexi- cal acquisition which jointly performs word segmentation and learns an explicit model of phonetic variation. We define the model as a Bayesian noisy channel; we sample segmen- tations and word forms simultaneously from the posterior, using beam sampling to control the size of the search space. Compared to a pipelined approach in which segmentation is performed first, our model is qualitatively more similar to human learners. On data with vari- able pronunciations, the pipelined approach learns to treat syllables or morphemes as words. In contrast, our joint model, like infant learners, tends to learn multiword collocations. We also conduct analyses of the phonetic variations that the model learns to accept and its patterns of word recognition errors, and relate these to de- velopmental evidence. 1 Introduction By the end of their first year, infants have acquired many of the basic elements of their native language. Their sensitivity to phonetic contrasts has become language-specific (Werker and Tees, 1984), and they have begun detecting words in fluent speech (Jusczyk and Aslin, 1995; Jusczyk et al., 1999) and learn- ing word meanings (Bergelson and Swingley, 2012). These developmental cooccurrences lead some re- searchers to propose that phonetic and word learning occur jointly, each one informing the other (Swingley, 2009; Feldman et al., 2013). Previous computational models capture some aspects of this joint learning problem, but typically simplify the problem consid- erably, either by assuming an unrealistic degree of phonetic regularity for word segmentation (Goldwa- ter et al., 2009) or assuming pre-segmented input for phonetic and lexical acquisition (Feldman et al., 2009; Feldman et al., in press; Elsner et al., 2012). This paper presents, to our knowledge, the first broad- coverage model that learns to segment phonetically variable input into words, while simultaneously learn- ing an explicit model of phonetic variation that allows it to cluster together segmented tokens with different phonetic realizations (e.g., [ju] and [jI]) into lexical items (/ju/). We base our model on the Bayesian word segmen- tation model of Goldwater et al. (2009) (henceforth GGJ), using a noisy-channel setup where phonetic variation is introduced by a finite-state transducer (Neubig et al., 2010; Elsner et al., 2012). This in- tegrated model allows us to examine how solving the word segmentation problem should affect infants’ strategies for learning about phonetic variability and how phonetic learning can allow word segmentation to proceed in ways that mimic the idealized input used in previous models. In particular, although the GGJ model achieves high segmentation accuracy on phonemic (non- variable) input …

Content truncated to 500 words