Baby SRL: Modeling Early Language Acquisition.

CoNLL 2008: Proceedings of the 12th Conference on Computational Natural Language Learning, pages 81–88 Manchester, August 2008 Baby SRL: Modeling Early Language Acquisition. Michael Connor Department of Computer Science University of Illinois connor2@uiuc.edu Yael Gertner Beckman Institute 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 A fundamental task in sentence compre- hension is to assign semantic roles to sen- tence constituents. The structure-mapping account proposes that children start with a shallow structural analysis of sentences: children treat the number of nouns in the sentence as a cue to its semantic predicate- argument structure, and represent language experience in an abstract format that per- mits rapid generalization to new verbs. In this paper, we tested the consequences of these representational assumptions via ex- periments with a system for automatic se- mantic role labeling (SRL), trained on a sample of child-directed speech. When the SRL was presented with representa- tions of sentence structure consisting sim- ply of an ordered set of nouns, it mim- icked experimental findings with toddlers, including a striking error found in children. Adding features representing the position of the verb increased accuracy and elim- inated the error. We show the SRL sys- tem can use incremental knowledge gain to switch from error-prone noun order fea- tures to a more accurate representation, demonstrating a possible mechanism for this process in child development. 1 Introduction How does the child get started in learning to in- terpret sentences? The structure-mapping view of early verb and syntax acquisition proposes that c⃝ 2008. Licensed under the Creative Commons Attribution-Noncommercial-Share Alike 3.0 Unported li- cense (http://creativecommons.org/licenses/by-nc-sa/3.0/). Some rights reserved. children start with a shallow structural analysis of sentences: children treat the number of nouns in the sentence as a cue to its semantic predicate- argument structure (Fisher, 1996), and represent language experience in an abstract format that per- mits rapid generalization to new verbs (Gertner et al., 2006). The structure-mapping account makes strong predictions. First, as soon as children can identify some nouns, they should interpret transitive and in- transitive sentences differently, simply by assign- ing a distinct semantic role to each noun in the sen- tence. Second, language-specific syntactic learn- ing should transfer rapidly to new verbs. Third, some striking errors of interpretation can occur. In “Fred and Ginger danced”, an intransitive verb is presented with two nouns. If children interpret any two-noun sentence as if it were transitive, they should be fooled into interpreting the order of two nouns in such conjoined-subject intransitive sen- tences as conveying agent-patient role information. Experiments with young children support these predictions. First, 21-month-olds use the number of nouns to understand sentences containing new verbs (Yuan et al., 2007). Second, 21-month-olds generalize what they have learned about English transitive word-order to sentences containing new verbs: Children who heard ”The girl is gorping the boy” interpreted the girl as an agent and the boy as a patient (Gertner et al., 2006). Third, …

Content truncated to 500 words