Exploring the Utility of Joint Morphological and Syntactic Learning from Child-directed Speech

Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 30–41, Seattle, Washington, USA, 18-21 October 2013. c⃝2013 Association for Computational Linguistics Exploring the utility of joint morphological and syntactic learning from child-directed speech Stella Frank sfrank@inf.ed.ac.uk Frank Keller keller@inf.ed.ac.uk ILCC, School of Informatics University of Edinburgh Edinburgh, EH8 9AB, UK Sharon Goldwater sgwater@inf.ed.ac.uk Abstract Children learn various levels of linguistic structure concurrently, yet most existing mod- els of language acquisition deal with only a single level of structure, implicitly assum- ing a sequential learning process. Developing models that learn multiple levels simultane- ously can provide important insights into how these levels might interact synergistically dur- ing learning. Here, we present a model that jointly induces syntactic categories and mor- phological segmentations by combining two well-known models for the individual tasks. We test on child-directed utterances in English and Spanish and compare to single-task base- lines. In the morphologically poorer language (English), the model improves morphological segmentation, while in the morphologically richer language (Spanish), it leads to better syntactic categorization. These results provide further evidence that joint learning is useful, but also suggest that the benefits may be dif- ferent for typologically different languages. 1 Introduction Models of language acquisition seek to infer lin- guistic structure from data with minimal amounts of prior knowledge, in order to discover which char- acteristics of the input data are useful for learn- ing, and thus potentially utilised by human learners. Most previous work has focused on learning individ- ual aspects of linguistic structure. However, children clearly learn multiple aspects in parallel, rather than sequentially, implying that models of language ac- quisition should also incorporate joint learning. Joint models investigate the interaction between different levels of linguistic structure during learning. These interactions are often (but not necessarily) synergis- tic, enabling better, more robust, learning by making use of cues from multiple sources. Recent models using joint learning to model language acquisition have spanned various domains including phonology, word segmentation, syntax and semantics (Feldman et al., 2009; Elsner et al., 2012; Doyle and Levy, 2013; Johnson, 2008; Kwiatkowski et al., 2012). In this paper we examine the joint learning of syntactic categories and morphology, which are ac- quired by children at roughly the same age (Clark, 2003b), implying possible interactions in the learn- ing process. Both morphology and word order de- pend on categorising words based on their morpho- syntactic function. However, previous models of syntactic category learning have relied principally on surrounding context, i.e., word order constraints, whereas models of morphology use word-internal cues. Our joint model integrates both sources of information, allowing the model to flexibly weigh them according to their utility. Languages differ in the richness of their mor- phology and strictness of word order. These char- acteristics appear to be (anti)correlated, with rich morphology co-occurring with free word order and vice versa (Blake, 2001; McFadden, 2003). The timecourse of acquisition is also influenced by lan- guage typology: learners of morphologically rich languages become productive in morphology ear- lier …

Content truncated to 500 words