Research and implementation on answer acquisition for question answering systems

Abstract

With the rapid development of the Web, people can easily store data, exchange information and share knowledge. Much knowledge which is used in daily life can be found on the Web. Nowadays, search engine has become the most important tool because it can help people to extract what they want from the Internet. However, users are limited to use several keywords to describe their requirement in search engines and can only obtain some related documents. Moreover, manually extracting the desired information from these related documents is a time-consuming job. Hence, Question Answering (QA) system which focuses on solving these problems has gradually attracted more and more researchers. In QA systems, users can use questions to describe their requirments and obtain answers to these questions which do not need to be refined. This is because, compared to the keywords used in search engines, these questions contain more semantic information to describe what the user wants more precisely, which enable QA systems return more accurate answers. QA systems can be categorized into automatic QA systems and user-interactive QA systems. The automatic QA systems use text matching or semantic matching to extract the answers, in which semantic information of the question target is first analyzed and then all the information which meets the requirement will be extracted from the document as the answers. The user-interactive QA systems rely on users offering the answers, in which questions are recommended to suitable users for answering. In this thesis, we focus on these two types of QA systems and research on the methods of improving the semantic analysis efficacy of questions, promoting the quality of related documents extracting, increasing the echo speed and recall of answers, and balancing the question recommendation mechanism. The main research areas and innovations of this thesis are as follows: Firstly, we propose a semantic pattern learning algorithm (SIIPU*S) in which an evaluation strategy named Semantic Identifiability Inverse Pattern Universality (SIIPU) is proposed to estimate the granularity of learned patterns for certain semantic requirement. In this algorithm, we study the relation between the syntactic constraints and semantic analyzing ability, and choose those suitable constraints to construct semantic patterns which can not only meet the requirement of semantic analysis but also cover more questions. Secondly, we utilize a query rewriting method in passage retrieval algorithm for extracting answer passages. In this algorithm, we use a heuristic query generation method to convert each question into some high quality queries, in which the weight of each keyword is determined by the corresponding question pattern. Therefore, those passages which hold the important words will be returned for answer extraction in advance. Thirdly, we propose a dynamic-pattern based answer extraction method, in which a heuristic rule learning method for information extraction which can automatically and efficiently acquire high-quality extraction rules from a user labeled training corpus. According to the semantic information of different questions, these rules can be dynamically converted into different types of answer patterns which can be used to precisely extract the answers to these questions from the related passages. Finally, we propose a balanced question recommendation method for user-interactive QA systems, in which system is responsiable for distributing each question to suitable users. In this algorithm, a user model is used to estimate the interests and professional areas of each user so that we can recommend questions to suitable users. To ensure most questions can be answered in time, a load balancing component is used to balance the work of each user and estimate the activity of each user to make sure of assigning emergent question to active users. Moreover, a question priority queue is maintained to ensure the emergent questions to be recommended earlier. On the basis of the above methods, we implement two QA prototype systems. Related experimental results show that these methods can improve the efficacy of QA systems effectively. IX Keywords: Web, Question Answer System, Question Recommendation, Information Extraction Rule, User Modeling, Semantic Pattern Learning