Identifying Lexical Relationships and Entailments with Distributional Semantics

As the field of Natural Language Processing has developed, research has progressed on ambitious semantic tasks like Recognizing Textual Entailment (RTE). Systems that approach these tasks may perform sophisticated inference between sentences, but often depend heavily on lexical resources like WordNet to provide critical information about relationships and entailments between lexical items. However, lexical resources are expensive to create and maintain, and are never fully comprehensive. Distributional Semantics has long provided a method to automatically induce meaning representations for lexical items from large corpora with little or no annotation efforts. The resulting representations are excellent as proxies of semantic similarity: words will have similar representations if their semantic meanings are similar. Yet, knowing two words are similar does not tell us their relationship or whether one entails the other. We present several models for identifying specific relationships and entailments from distributional representations of lexical semantics. Broadly, this work falls into two distinct but related areas: the first predicts specific ontology relations and entailment decisions between lexical items devoid of context; and the second predicts specific lexical paraphrases in complete sentences. We provide insight and analysis of how and why our models are able to generalize to novel lexical items and improve upon prior work. We propose several shortand long-term extensions to our work. In the short term, we propose applying one of our hypernymy-detection models to other relationships and evaluating our more recent work in an end-to-end RTE system. In the long-term, we propose adding consistency constraints to our lexical relationship prediction, better integration of context into our lexical paraphrase model, and new distributional models for improving word representations.