Monitoring distributional patterns in our input: probabilistic models for production and comprehension

old_uid10150
titleMonitoring distributional patterns in our input: probabilistic models for production and comprehension
start_date2011/09/09
schedule13h30-15h
onlineno
summaryIn this talk I will propose two different probabilistic models that show how our language production and comprehension heavily rely on keeping track of patterns present in what we hear. The first part of the talk concentrates on children's production. Focusing on the dative alternation in English, I examine whether children's choices are influenced by the same factors that influence adults' choices. I also look at whether children are sensitive to multiple factors simultaneously. Using mixed effects regression models, I find parallels between child and adult speech, consistent with recent acquisition research suggesting that there is a usage-based continuity between child and adult grammars. The results demonstrate that from early on children pay attention to complex distributional patterns, and replicate the subtle patterns found in their input. The second part of the talk targets adults' comprehension when dealing with errorful input. I examine number agreement errors, which are relatively common in production. I suggest that comprehension of such input reflects an understanding of likely production errors: comprehenders know the likelihood of different kinds of production errors, and recover more easily from more probable errors. I use a comprehension model which represents an optimal allocation of processing resources under noisy input (Levy 2008). The results show that language-specific estimates of these likelihoods result in a better fit to comprehension data, and that such a model predicts behavior in two domains: syntactically well-formed local coherences as reported by Tabor et al. (2004) and syntactically ill-formed agreement errors as reported by Pearlmutter et al. (1999).
responsiblesCandito