Semantics driven anaphora resolution
Håvar Skaugen · NORA - Norwegian Open Research Archives · 2015
This thesis describes a method for generating semantically motivated antecedent candidates for use in pronominal anaphora resolution. Predicate-argument structures are extracted from a large corpus of text parsed by the NorGram grammar and used as the basis for a fuzzy classification model. Given a pronominal anaphor, the model generates antecedent candidates ranked by the frequency by which they co-occur in the same lexical context as the anaphor. This set of candidates is intersected with the set of nouns gathered from the anaphor's recent context. A selection basic heuristics are then introduced to the model in a permutational fashion to gauge their individual and combined effect on the model's accuracy. The model reached an accuracy of 56.22% correct predictions. Additionally, in a slightly modified model the correct antecedent was found within the antecedent candidate list for 87.12% of the anaphora.