Abstract

Textual entailment (TE) is a relation that holds between two pieces of text where one reading the first piece can conclude that the second is most likely true. Accurate approaches for textual entailment can be beneficial to various natural language processing (NLP) applications such as: question answering, information extraction, summarization, and even machine translation. For this reason, research on textual entailment has attracted a significant amount of attention in recent years. A robust logical-based meaning representation of text is very hard to build, therefore the majority of textual entailment approaches rely on syntactic methods or shallow semantic alternatives. In addition, approaches that do use a logical-based meaning representation, require a large knowledge base of axioms and inference rules that are rarely available. The goal of this thesis is to design an efficient description logic based approach for recognizing textual entailment that uses semantic relatedness information as an alternative to large knowledge base of axioms and inference rules.

In this thesis, we propose a description logic and semantic relatedness approach to textual entailment, where the type of semantic relatedness axioms employed in aligning the description logic representations are used as indicators of textual entailment. In our approach, the text and the hypothesis are first represented in description logic. The representations are enriched with additional semantic knowledge acquired by using the web as a corpus.
The hypothesis is then merged into the text representation by learning semantic relatedness axioms on demand and a reasoner is then used to reason over the aligned representation. Finally, the types of axioms employed by the reasoner are used to learn if the text entails the hypothesis or not. To validate our approach we have implemented an RTE system named AORTE, and evaluated its performance on recognizing textual entailment using the fourth recognizing textual entailment challenge. Our approach achieved an accuracy of 68.8 on the two way task and 61.6 on the three way task which ranked the approach as 2nd when compared to the other participating runs in the same challenge. These results show that our description logical based approach can effectively be used to recognize textual entailment.