Semantic pattern for question answering system

Abstract

With the dramatic development of the Internet and the emergence of Web 2.0, User-Interactive Question Answering (UIQA) systems have been developed and become very popular Web-based services. Unlike the traditional automatic Question Answering (QA) systems which obtain answers automatically, the User-Interactive QA systems serve as interactive platforms for users to help each other with human-provided answers, which overcome the shortcoming of poor quality of the automatic answers. Surface pattern is proved an effective way to retrieve answers automatically. However, surface pattern does not include semantic information and is therefore called "poor-knowledge approaches". Hence, it cannot extract precise answers or relevant information without semantically analyzing questions and answers. To solve this problem, we firstly propose a novel type of pattern called semantic pattern and give the formal definition. The architecture of UIQA system based on semantic pattern is also presented, which includes question structure analysis, pattern matching, pattern generation, pattern classification and answer extraction. After that, to generate semantic pattern automatically and effectively, this thesis proposes a new automatic generation method of semantic patterns from free-text questions. This method uses structural processing and name entity recognition (NER) to obtain the main structure of a question. An entropy-based model is used to select suitable words from questions for generalization. WordNet is then applied in our algorithm to get the best semantic labels from our Tagger Ontology for such chosen words. An evaluation method is also proposed to estimate the suitability of the generated patterns and is implemented in a real UIQA system. Experiments with 5500 questions show that 63.9% generated patterns are satisfactory on average. Finally, this thesis presents one of the applications of semantic pattern as an example - an automatic method for building a semantic dictionary from existing semantic pattern based questions for question categorization. This dictionary consists of two main parts: Semantic Domain Terms (SDT), which is a domain specific term list, and Semantic Labeled Terms (SLT), which contains common terms tagged with semantic labels. We implement the semantic dictionary construction method on a set of 2509 questions with semantic patterns in our system. Experimental results show that the precision of question classification is improved by 7.5% on average after using the constructed semantic dictionary compared with the baseline method.