Research On Question And Answer Matching Oriented Online Shopping

Online shopping has become a necessary in people’s lives. It is convenientand fast, allowing users to browse and shopping in home. However, onlineshopping customer service often chat with multiple user and reply a lot questionsat the same time, resulting in the loss of customers. If there is an assistingquestion answering system help customer retrieve information and gives somesuggestions for the answer, it will improve the efficiency and quality of servicecustomer service greatly. Question answering system has wide application inonline shopping consultation.The performance of a question answering system depends on the knowledgebase’s quantity and scale. So, it has become a core problem extracting from chatlog. There are a number of questions corresponding with a number of answers inthe corpus of network session. The most obviously feature of the complexcorresponding is the answer lag. Buyers raised a number of questions, customersservice answer them one by one. Questions and answers may not be adjacent. Atpresent, most of the question answering system knowledge base of question-answer pairs are artificial constructed, which not only is time-consuming andhigh maintenance cost, but also cannot update in real time.To solve this problem, we propose a self-training model framework thatextract question answering pairs from a large chat records. The framework canuse a small labeled corpus and a mount unlabeled corpus to iterate training model.In this framework includes three different models: the feature detection basedmatching evidences model which include sentence type, common word sequence,the concept of the relationship; the redundant information based relation modelwhich compute similarity between the searching relevance answers and thecandidate answer; the word co-occurrence based relation model which statist thecommon vocabulary sequence between question and answer.In the corresponding experiments, test each model and analysis their results.Three models are effective for extracting question-answer pairs from session record. Finally combine three models together, forming a self-training framework.The accuracy of the framework is greatly improved after the combination.