Extractive Answer Fusion For Community Question Answering

Community-based question answering provides a platform for users,which allows them to ask question and get the answers they need easily.It makes full use of human-computer interaction,forms a large-scale user generated content corpus.However,these QA pairs have two major problem: a)CQA reflects the language habits of people's natural state,and the redundancy provided by different users is relatively high;b)Maybe all users can provide some useful information,but none can cover all aspects.CQA contains a large number of viewpoint-descriptive questions.The answer to such questions is characterized by weak exclusivity.Therefore,it is necessary to reintegrate multiple QA pairs that contain different answers to the same question through the answer fusion method.In order to give the user a concise and comprehensive answer,this paper divides the answer fusion task into answer selection and sentence matching.The main content of this paper includes the following four aspects:(1)This paper treats answer selection as a classification problem in chapter 2,and uses supervised methods to extract the lexical features,syntactic features,and shallow semantic features to training.(2)This paper treats answer selection as a ranking problem in chapter 3,and uses neural networks to train the word embedding to extract deep semantic features and logical features.This paper tries different sentence coding structures on CNN and LSTM models to conduct comparative experiments.Integrating attention mechanisms into answer selection tasks.This dissertation implements inner-attention mechanism and attention-over-attention mechanism on answer selection model.Attention mechanism can reasonably allocate the weights of different words,and obtain a more accurate sentence vector representation during encoding layer.The experimental results show that the model based on AOA and LSTM answer selection model obtains the best scores.(3)Calculating sentence similarity in several different ways and improving the sentence-matching algorithm by mixing these methods in chapter 4.Then this paper constructs a similarity matrix to realize the sentence-matching model.Finally,this paper implements a sematic alignment model based on LSTM.It extracts features from word-level and sentence-level,which increases the accuracy.(4)Designing and implementing an automated answer fusion system in chapter 5.This system can retrieve relevant questions and answers from the question entered by the user,then synthesize multiple documents to extract specific answers that can answer the entered question and compare the similarities of the answers.Finally,this system shows the fusion result to the users,which can improve the feedback experience.