Research On Learning To Rank Answers For Vcsa

With the rapid development of Internet and the express industries, online shoppinghas become an indispensable shopping mode for most people. Online shopping is soconvenient and efficient that people could buy things at home instead of in markets.Sellers usually offer online customer services to communicate with buyers. For the sakeof lower cost, one customer-service staff usually deals with two or three buyers at thesame time. As a consequence, the quality of customer service decreased. So peoplecome up with an idea of building a Virtual Conversational Sales Assistant (VCSA) toanswer buyer’s questions by chat logs.The performance of a VCSA is greatly influenced by answer ranking. Currentanswer ranking strategies of VCSA are usually based on single relevance of twosentences. But these methods cannot achieve fine ranking performance because theysimply considered only one aspect of relevance by ignoring others like syntactic andsemantic level features. If the answers from VCSA are not precise enough, the workefficiency of customer-service staff will get down. So it is urgent that improving theanswer ranking performance of VCSA.In order to solve these problems, we used a learning to rank method to deal withthe answer ranking task by transferring ranking problems to classification issues. Somespecific feature extraction approaches are proposed for the method. There are two traitsof the online shopping corpus: the first one is that the sentences are usually brief andoral and they are short text; the second one is that every sentence has its context. So wedesigned nine features of similarity calculation for the ranking on the semantic andsyntactic levels. The semantic features are: the semantic relevance calculations based onHowNet, Wikipedia, Synonyms Dictionary and words co-occurrence method; thesyntactic features are: the similarities of sentences based on TF/IDF, BM25and the ratioof word numbers. Then we extended the features by using the context information. Atlast, a feedback-based ranking model is proposed to improve the performance of VCSA.In order to show the improvements of these two methods, we implemented aVCSA system and designed two experiments for comparison: the first experimentcompares the ranking performance by the one-relevance similarity calculation with themethod of learning to rank. The result shows that the latter is better than the former. Thesecond experiment compares the ranking performance between with and without thefeature extension by context and it turns out that feature extension by context is veryhelpful to improve the ranking performance of VCSA.