Research On Crucial Technology Of Q&A System In Endowment Insurance

The endowment insurance is closely related to people’s livelihood, the majority of theinsured masses in the enrollment process often encounter a variety of problems. Theresearch on key technologies of question answering system and its application to the field ofpension insurance, which can answer users’ questions quickly and accurately with naturallanguage, has profound research and practical significance.This thesis does research on Question-answering technology in the field of pensioninsurance, and mainly complete works as the following: Firstly, we analyze thecharacteristics of140,000real questions related to topics in pension insurance field fromwebsites, and classified to different categories by domain topics. We also combine withdomain thesaurus in questions understanding stage for word segmentation, semanticannotation, keyword extraction, recognition of named entity and analysis of syntacticdependency. Secondly, we put forward an applicable question classification algorithmnamed KNN active learning for pension insurance field. It extracts a small part from thecollection of Q&A in pension insurance field to annotate manually, and uses heuristicmethod to mainly select data more beneficial for classification model and annotate it, extendthe sample set that has been marked, reduce the sample set to be marked. It learns iterativelybased on this, which conduces to higher classification accuracy of the new model bytraining. The question that the question classifier has low accuracy under smaller sample setof annotation is solved by this method.Thirdly, we give a new question retrieval modelnamed MFISC. It uses weighted arithmetic idea to compute similarity value from lexicalsimilarity, similarity of question vocabulary’s semantic meaning, and similarity of syntacticstructure. The result obtained is used as similarity between query input and historicalquestions, to improve the retrieval accuracy. Ultimately the answer from historical questionthat has the highest similarity will be returned as answers for current query question.Lastly, we complete the experiments of the above two algorithms to prove the validityof question classification algorithm based on KNN active learning and the retrieval modelnamed MFISC. The test results show that our work has practical application value forconstruction of Q&A system in the field of pension insurance.