Title: Psycholinguistically-motivated partial supervision for grammar induction
Grammar induction, the discovery of syntactic structure from raw text, is not only useful for downstream NLP tasks such as question answering and machine translation. It is also interesting as computational modeling of language acquisition in humans. Unlike the standard grammar induction evaluation setting, in which no prior knowledge of input is assumed, child and adult language learners have access to limited kinds of lexical knowledge, such as knowledge of certain common nouns (Shi et al., 1999; Waxman and Booth, 2001; Bergelson and Swingley, 2012; Waxman et al., 2013) or function words (Selkirk, 1984, 2008; Christophe et al., 2008), and this type of information is thought to constrain syntactic acquisition (Gleitman, 1990; Naigles, 1990; Gertner et al., 2006; Peterson Hicks, 2006; Fisher et al., 2010; Jin and Fisher, 2014).
In this talk, I describe experiments using an existing memory-bounded left-corner parsing model of grammar induction (Shain et al., 2016), augmented with a small number of common word types of frequently occurring categories. Results show substantial improvement for nouns in particular over a baseline with no augmentation, and over the results of similar augmentation with other closed- and open-class word types.