Word Semantic Representations using Bayesian Probabilistic Tensor Factorization

Many forms of word relatedness have been developed, providing different perspectives on word similarity. We introduce a Bayesian probabilistic tensor factorization model for synthesizing a single word vector representation and per-perspective linear transformations from any number of word similarity matrices. The resulting word vectors, when combined with the per-perspective linear transformation, approximately recreate while also regularizing and generalizing, each word similarity perspective. Our method can combine manually created semantic resources with neural word embeddings to separate synonyms and antonyms, and is capable of generalizing to words outside the vocabulary of any particular perspective. We evaluated the word embeddings with GRE antonym questions, the result achieves the state-ofthe-art performance.