Predicting Processing Cost of Anaphora Resolution

International audienceAnaphora resolution, for human speakers, can be more or less costly depending on various factors like ambiguity, syntactic complexity and semantic plausibility. The variation of cost has been measured by many studies in psycholinguistics, through experimental paradigms like self-paced reading, or eye-tracking. Our project aims at devising a system, inspired by current NLP coreference resolution systems, that can predict a processing cost for anaphora resolution, which can be evaluated by running our system on human data coming from psycholinguistic experiments, or eye-tracking corpora e.g. the Dundee Corpus (Kennedy 2003)