ComText: Unsupervised Word-Sense Disambiguation Through Statistical Network Analysis and Typed Dependencie

Abstract

The task of determining the context of different words and specific senses is a crucial aspect of natural language processing, for a machine that cannot ascertain the correct meaning of the words of an input cannot glean any further form of understanding. In this work, we propose ComText - an algorithm for statistical, network-based word-sense disambiguation that utilizes a multi-layer network to generate similar words and topics based on relative proximity. We build upon the Stanford natural language parser and previous work on typed dependencies to determine the syntax from the phrase structures of our training corpus, and use these dependencies to create network maps that are then used to analyze the textual input.