Research On The Question Answering System For The Online Call Center

Question Answering System is an important branch in the study of Natural Language Pro-cessing. It aims at allowing people to have answers when using natural language to ask. This Question Answering System, based on sets of frequently asked questions, could give more ac-curate and effective answers. For a long time, how to acquire a massive sets of frequently asked questions with high quality is always the developmental bottleneck of such system. However, the existence of thousands of data of human services on internet, produced by those human services staff, provides a new development opportunity for this system.Based on 230 thousand human service materials and using automatic human service plat-form as practical background, several work has been accomplished:At first, a complete question answering base-system based on sets of frequently asked questions is designed. In the frame of this system, by fully and comprehensively using Chinese Word Segmentation、Named Entity Recognition、Chinese Part-of-Speech Tagging、Parsing、 Keyword Extraction、Information Retrieval、Similarity Calculation and other similar technol-ogy and algorithms, a complete solution for establishing question answering system based on sets of frequently asked questions of massive data is acquired.Secondly, the critical module of this system, similarity calculation of questions, is mainly studied. By comparing several traditional similarity calculation algorithm and combining the actual practical background, the flaws of using traditional similarity calculation algorithm to solve problems studied in this article are pointed out. On the basis of research and analysis, a more detailed improved solution is achieved by conducting research in different angles such as weight tuning、Synonym expansion、word alignment、dependency parsing. Besides, by intro-ducing liner model and using parking algorithms to combine these similarity calculation algo-rithms, the performance of this system is furthermore improved.At last, the experiments shows the analysis of the theory is correct and the improved algo-rithm is effective. And for the similarity calculation algorithm with worse performance, a fur-thermore fault analysis is conducted and an improved solution is acquired.A complete system framework based on frequently asked questions is provided in this article, with emphasis research on similarity calculation. There are some progresses, however, also with some drawbacks. For example, in synonym expansion the existence of a general phe-nomenon, one word with different meanings, brings new noise for similarity calculation, thereby a further Word sense disambiguation should be conducted. Besides, the utilization of dependency parsing will also significantly decreases the respond time of the system.