Research On Candidate Answer Sentence Extraction For Baidubaike

Comparing to traditional search engine,question answering system can analyze users' questions more generally,can locate users' document and answers more precisely.Thus,question answering system,which core is locating the document and answers,become the one of hot points of Natural Language Processing area,also the one of unsolved problems.According to different corpus,location of answers has different technical proposal and research area.But today,the big data age,the structured or semi-structured corpus cannot cover all kinds of QA,which causes the arising of answer location technology for natural-language-formed document.The point of this paper is the answer extraction for natural-languageformed document.We proposed three methods of answer extraction: answer extraction based on word match,answer extraction based on traditional machine learning and answer extraction based on deep learning.Besides we proposed a method that converged multi technology.Answer extraction based on semantic matching has been proposed many years.In this paper,we use word-overlap to compute the similarity between question and answers,also use the sentence similarity measure method bas ed on word similarity,which has two ways to get,including method based on word embedding and method based on How Net.Machine learning method need us analyze the relationship between question and answers manually to find some useful feature.Then use support vector machine to get scores of each answer for the final answer selection.Our experiment shows that the selecting of feature has the most importance of the answer extraction task.Deep learning method can learn the feature automatically which avoid the great amount of feature engineering job.We use multi deep learning structure to do the experiment,and the results show that the attention-based GRU model can learn the representation of sentence better and got the best result.Finally,we proposed a converged method that combine the previou s experiment.We converged multi model together,making them get best use so that easy model can handle easy problem,complex model can handle complex problem,which got the best experiment result.