Graph-based Algorithms for Lexical Semantics and its Applications

Abstract

Lexical semantics studies the meaning of words, which is a useful tool for computer-based automatic natural language processing (NLP). This thesis explores graph-based algorithms to learn and apply distributional lexical semantics in NLP applications. One theory of lexical semanticists holds that semantic relations among words can be extracted from their textual context in natural languages. Based on this theory, we propose using graphs to represent natural language text according to the contextual relations of words in higher-level language units (e.g. sentences, definitions or documents). In these graphs, words and/or higher-level language units are represented with nodes, and edges are added between them according to their textual context to indicate their observed relatedness in a dictionary or a collection of documents. We explore two types of graph representations: the word-word graph which is used for modeling the semantic relations among words, and the instance-word bipartite graph which uses words as a medium to study the relatedness among higher-level language units. In this way, we can embed the semantic relations among words and optionally higher-level language units into the graph structure. We design algorithms to propagate semantic information through the graphs in order to recover the unobserved relatedness among words or higher-level language units, which is used in designing unsupervised, semi-supervised or active learning algorithms to reduce human supervision in NLP applications for harvesting or analyzing text data resources. Specifically, we design graph-based algorithms either for quantitatively assessing lexical semantic similarity, or for developing representativeness and diversity criteria for selecting a characteristic subset of terms, which is useful for problems such as keyword summarization and query design for active learning. In particular, we focus on designing graph-based algorithms to apply lexical semantics in three NLP applications, including Wiktionary lexical semantic similarity extraction, Twitter user interest extraction, and active learning for semantic orientation classification.