Generating and applying textual entailment graphs for relation extraction and email categorization

Abstract

Recognizing that the meaning of one text expression is semantically related to the meaning of another can be of help in many natural language processing applications. One semantic relationship between two text expressions is captured by the textual entailment paradigm, which is defined as a relation between exactly two text expressions. Entailment relations holding among a set of more than two text expressions can be captured in the form of a hierarchical knowledge structure referred to as entailment graphs. Despite the fact that several people have worked on building entailment graphs for different types of textual expressions, little research has been carried out regarding the applicability of such entailment graphs in NLP applications. This thesis fills this research gap by investigating how entailment graphs can be generated and used for addressing two specific NLP tasks: First, the task of validating automatically derived relation extraction patterns and, second, the task of automatically categorizing German customer emails. After laying a theoretical foundation, the research problem is approached in an empirical way, i.e., by drawing conclusions from analyzing, processing, and experimenting with specific task-related datasets. The experimental results show that both tasks can benefit from the integration of semantic knowledge, as expressed by entailment graphs.