Acquiring language from speech by learning to remember and predict

Proceedings of the 24th Conference on Computational Natural Language Learning, pages 195–214 Online, November 19-20, 2020. c⃝2020 Association for Computational Linguistics https://doi.org/10.18653/v1/P17 195 Acquiring language from speech by learning to remember and predict Cory Shain Ohio State University shain.3@osu.edu Micha Elsner Ohio State University elsner.14@osu.edu Abstract Classical accounts of child language learning invoke memory limits as a pressure to dis- cover sparse, language-like representations of speech, while more recent proposals stress the importance of prediction for language learn- ing. In this study, we propose a broad- coverage unsupervised neural network model to test memory and prediction as sources of signal by which children might acquire lan- guage directly from the perceptual stream. Our model embodies several likely properties of real-time human cognition: it is strictly incre- mental, it encodes speech into hierarchically organized labeled segments, it allows interac- tive top-down and bottom-up information flow, it attempts to model its own sequence of la- tent representations, and its objective function only recruits local signals that are plausibly supported by human working memory capac- ity. We show that much phonemic structure is learnable from unlabeled speech on the basis of these local signals. We further show that re- membering the past and predicting the future both contribute to the linguistic content of ac- quired representations, and that these contribu- tions are at least partially complementary. 1 Introduction How children acquire language from the envi- ronment is one of the fundamental mysteries of cognitive science. Much theoretical, experimen- tal, and computational research into this question has focused on acquiring abstractions over lower- order symbols, such acquiring morphemes from phoneme sequences or syntactic structures from word sequences (Chomsky, 1965; Gold, 1967; El- man, 1991; Saffran et al., 1996; Albright, 2002; Klein and Manning, 2004; Goldwater et al., 2009; Christodoulopoulos et al., 2012, inter alia). Chil- dren, however, do not get symbolic input; sym- bolic representations at any level of granularity constitute abstractions inferred from highly vari- able, noisy, and information-rich perceptual signals like audition and vision. This work joins a grow- ing computational literature exploring the kinds of architectures and learning objectives that best support acquisition of linguistic representations di- rectly from the speech signal without supervision (Versteegh et al., 2015; Dunbar et al., 2017). Such models can be used to test questions about lan- guage acquisition under more realistic assumptions about the input signal, especially to the extent that they reflect known constraints on human cognition (Shain and Elsner, 2019; Beguˇs, 2020). This study uses computational modeling to ex- amine two influential and possibly complementary ideas about how people learn abstract representa- tions, including language, from data: learning to remember, and learning to predict. Both hypothe- ses have been advocated by prior work in language acquisition, cognitive neuroscience, and computa- tional modeling, yet their relative contributions to language learning are not yet clear. Our model per- mits precise manipulation of memory and predic- tion pressures during acquisition, allowing direct comparison of these hypotheses. In so doing, we implement several constraints on real-time language …

Content truncated to 500 words