Minimally Supervised Model of Early Language Acquisition

Proceedings of the Thirteenth Conference on Computational Natural Language Learning (CoNLL), pages 84–92, Boulder, Colorado, June 2009. c⃝2009 Association for Computational Linguistics Minimally Supervised Model of Early Language Acquisition Michael Connor Department of Computer Science University of Illinois connor2@uiuc.edu Yael Gertner Department of Psychology University of Illinois ygertner@cyrus.psych.uiuc.edu Cynthia Fisher Department of Psychology University of Illinois cfisher@cyrus.psych.uiuc.edu Dan Roth Department of Computer Science University of Illinois danr@uiuc.edu Abstract Theories of human language acquisition as- sume that learning to understand sentences is a partially-supervised task (at best). Instead of using ‘gold-standard’ feedback, we train a simplified “Baby” Semantic Role Labeling system by combining world knowledge and simple grammatical constraints to form a po- tentially noisy training signal. This combina- tion of knowledge sources is vital for learn- ing; a training signal derived from a single component leads the learner astray. When this largely unsupervised training approach is ap- plied to a corpus of child directed speech, the BabySRL learns shallow structural cues that allow it to mimic striking behaviors found in experiments with children and begin to cor- rectly identify agents in a sentence. 1 Introduction Sentence comprehension involves assigning seman- tic roles to sentence constituents, determining who does what to whom. How do young children be- gin learning to interpret sentences? The structure- mapping view of early verb and syntax acquisition proposes that children treat the number of nouns in the sentence as a cue to its semantic predicate- argument structure (Fisher, 1996), and represent lan- guage experience in an abstract format that promotes generalization to new verbs (Gertner et al., 2006). Theories of human language acquisition assume that learning to understand sentences is naturally a partially-supervised task: the fit of the learner’s predicted meaning with the referential context and background knowledge provides corrective feed- back (e.g., Pinker (1989)). But this feedback must be noisy; referential scenes provide ambiguous in- formation about the semantic roles of sentence par- ticipants. For example, the same participant could be construed as an agent who ’fled’ or as a patient who is ’chased’. In this paper, we address this problem by de- signing a Semantic Role Labeling system (SRL), equipped with shallow representations of sentence structure motivated by the structure-mapping ac- count, that learns with no gold-standard feedback at all. Instead, the SRL provides its own internally- generated feedback based on a combination of world knowledge and linguistic constraints. As a sim- ple stand-in for world knowledge, we assume that the learner has animacy information for some set of nouns, and uses this knowledge to determine their likely roles. In terms of linguistic constraints, the learner uses simple knowledge about the possible ar- guments verbs can appear with. This approach has two goals. The first is to in- form theories of language learning by investigating the utility of the proposed internally-generated feed- back as one component of the human learner’s tools. Second, from an NLP and Machine Learning per- spective we propose to inject information into a su- pervised learning algorithm through a channel …

Content truncated to 500 words