Learning Embeddings for Transitive Verb Disambiguation by Implicit Tensor Factorization

We present an implicit tensor factorization method for learning the embeddings of transitive verb phrases. Unlike the implicit matrix factorization methods recently proposed for learning word embeddings, our method directly models the interaction between predicates and their two arguments, and learns verb phrase embeddings. By representing transitive verbs as matrices, our method captures multiple meanings of transitive verbs and disambiguates them taking their arguments into account. We evaluate our method on a widely-used verb disambiguation task and three phrase similarity tasks. On the disambiguation task, our method outperforms previous state-ofthe-art methods. Our experimental results also show that adjuncts provide useful information in learning the meanings of verb phrases.