Research On The Questions Similarity In The FAQ Answering System

FAQ module is an important part of answering system. FAQ is formed throughstoring common questions and their answers, which can help improving theefficiency of answer seeking and synthesizing for similar questions. When usingFAQ in answering system, there are two key problems, including question-updating and question-matching for new problems. And the key to these problemsis the method of question similarity computation.We focus on the questions similarity computation in FAQ system. We aim tofind a new questions similarity computation method, which can use thecharacteristics of Chinese and Chinese questions to improve the Chinese questionssimilarity accuracy, and meanwhile to improve the performance of the Chinesequestion answering system.Our contributions are as follows:(1) In the existing literatures, questions similarity computation methods aremainly based on general declarative sentence similarity computation methods,where only semantic similarity between two general declarative sentences isconsidered. While considering questions similarity, both similarity between twoquestions and similarity between two answers should be considered. So a newquestions similarity computation method is proposed, which takes into account notonly semantic and syntactic characteristics between questions, but also informationsuch as question type to indirectly get the characteristics of the answer, in order toextract deep semantic information of question, and finally to improve questionssimilarity computation accuracy. Experimental results show that the method canachieve better accuracy.(2) For question similarity computation method based on Hownet, same wordsimilarity method is adopted for those specific words such as contrastive words,contradictive words and synonymous words, which might result that two oppositequestions have wrong high similarity. For all this, an improved questions similaritycomputing method based on Hownet is proposed. In this method, not onlyorientation inference between Chinese words but also orientation inference betweenChinese word and English word can be computed correctly. And, contrastive orcontradictive words have a negative effect on questions similarity computation, which reduces the possibility that opposite questions have high similarity.Empirical results show the validity of the method.(3) A prototype of FAQ answering system is designed to provide some relatedalgorithms from our research, and a platform for future in-depth studies.