Acquiring Verb Argument Structure from the Input Distribution: An Artificial Language Study

Add to your list(s) Download to your calendar using vCal

• Dr Elizabeth Wonnacott, University of Oxford, Department of Experimental Psychology
• Tuesday 16 October 2007, 16:00-17:30
• GR-06/07, English Faculty Building.

If you have a question about this talk, please contact Napoleon Katsos.

Adult language combines a complex mix of regular, rule-like processes and more conservative, lexically based patterns. For example, verb argument structures may generalize to new verbs (Arthur flupped the ball to Ford->Arthur flupped Ford the ball) yet resist generalization with certain lexical items (Arthur carried the ball to Ford → *Arthur carried Ford the ball). Pinker (1989) suggests that that learning whether a particular verb occurs with a particular argument structure involves learning a fine-grained semantic representation. However, frequency-based entrenchment effects in young children (Theakston 2004) and statistical effects in sentence processing (Trueswell et al 1993), suggest that learners track verb-structure co-occurrences. This concurs with recent approaches which emphasize the role of Statistical Learning processes in Language Acquisition (Saffran et al. 1996). Our work uses Artificial Language Learning to explore whether the relationship between verbs and argument structures can be acquired from input statistics. Adult participants were exposed to languages in which input distribution provided the only cue to verb subcategory (no semantic or phonological correlates). Each language had 12 verbs and two synonymous argument structures, VSO and VOS -Particle. Four languages were explored, with differences among them in the degree to which their verbs exhibited lexically based versus language-wide patterns. After several days of exposure, learners took Production, Grammaticality Judgment and Eye-tracked Comprehension tests. In this last, participants heard sentences and viewed two scenes: correct and agent-patient reversed. If the verb is biased to occur with one structure, eye-movement data should reveal a looking preference before the disambiguating particle.

Overall, participants proved able both to acquire verb subcategorization, and to generalize. Behavior in the different tests was very consistent, and reflected both the statistical preferences of particular verbs and across-verb statistics (shown with ‘new’ verbs introduced during testing and not phonologically/semantically related to ‘old’ verbs). Eye-movement data revealed that these statistics influenced on-line processing. However, the tendency to generalize was affected by a third source of information: the distribution of verb types in the language. Learners exposed to languages in which the majority of verbs occurred in multiple structures were more likely to generalize.

In conclusion, learning and processing of verb argument structure is strongly driven by various distributional properties of the input, and these purely formal phenomena may occur in the absence of semantic cues. In addition, Artificial Language learning is shown to provide a fruitful methodology for exploring the relationship between Statistical Learning and statistical effects in sentence processing.

This talk is part of the RCEAL Tuesday Colloquia series.

This talk is included in these lists:

• All Talks (aka the CURE list)
• GR-06/07, English Faculty Building
• RCEAL Tuesday Colloquia

Note that ex-directory lists are not shown.