A predictive model for human parsing
Speaker: Vera Demberg
Institution: University of Edinburgh
Abstract:
In this talk I present ideas and preliminary results for a predictive model of human parsing. The goal of our work is to predict difficulties in human sentence processing by making predictions about how a sentence should be continued. Predictions are computed based on data from large corpora. If predictions are verified in the course of the sentence, we expect shorter processing time, whereas processing is expected to be delayed when a surprising (i.e. unexpected) constituent is encountered.
We here report results from a pilot study on prediction in the processing of relative clauses. We extracted different types of relative clauses from an English eye-tracking corpus. Subject relative clauses (SRC, e.g. "The dog that saw me.") have a higher prior probability than object relative clauses (ORC, e.g. "The dog that I saw."). Therefore, we expect reading times to be longer on ORCs (surprisal to see a article or other part of a noun phrase instead of an adverb or verb) than on SRCs. We computed a linear regression model for the total reading times of the relative pronoun and the spill-over region (the first words of the relative clause) in dependence of word length, POS tag, subjects, the relative pronoun, word frequency, and type of relative clause. In this model, we found the type of the relative clause (ORC vs. SRC) to correlate with total reading time in the corpus, which provided preliminary evidence for a predictive model of relative clause processing that takes prior probability into account.