A cognitively plausible model for anaphora resolution

Anaphora resolution a is complex problem, as it deals with syntax, semantics and discourse. The subject is well studied in field of psycholinguistics, where multiple preferences were discovered, and in the field of computational linguistics, where many systems have been developed to perform anaphora resolution in documents. Nevertheless, the work done in the two fields remains disconnected. 

We investigate how we can bridge the gap by exploiting the options for making a cognitively plausible model for anaphora resolution. Such model can be beneficent for both fields as it can inspire computational linguistics with findings about how humans process anaphora, and help the psycholinguistic community developing large coverage, incremental models simulating the human processing of anaphora. 

Inspired by the surprisal framework that uses incremental probabilistic parsing to predict processing cost coming from syntax for each word in a corpus, we focus on incremental probabilistic systems of anaphora resolution from the field of computational linguistics, for example the largely spread pair-wise model, and turn them into a cognitive model of processing of anaphora. We will present some preliminary results on the measures of processing cost of anaphora we developed and the perspective of evaluating the model on eye-tracking data.