Non-negative Matrix Factorisation-Based Verb Semantics for 3rd Person Pronoun Resolution

We present an initial utility study of a distributionalmodel of verb selectional preferences for 3rd personpronoun resolution in German. We investigate cases in which3rd person pronouns occur as subjects of transitive verbs. Ineach such case, the likelihood of inserting one of the antecedentcandidates is calculated as the conditional probability of theantecedent candidate given either the verb governing thepronoun or the object of the verb. These probabilities areestimated using a matrix derived from frequency counts ina large corpus. Non-negative matrix factorisation is appliedas a sort of semantic smoothing to address the sparsity issueinherent in the approach.