Semantic Information Extraction for Improved Word Embeddings

Word embeddings have recently proven useful in a number of different applications that deal with natural language. Such embeddings succinctly reflect semantic similarities between words based on their sentence-internal contexts in large corpora. In this paper, we show that information extraction techniques provide valuable additional evidence of semantic relationships that can be exploited when producing word embeddings. We propose a joint model to train word embeddings both on regular context information and on more explicit semantic extractions. The word vectors obtained from such an augmented joint training show improved results on word similarity tasks, suggesting that they can be useful in applications that involve word meanings.