Research On Knowledge-based Question Answering

A typical knowledge-based question answering (KB-QA) system faces twochallenges: one is to transform natural language questions into their meaningrepresentations (MRs); the other is to retrieve answers from knowledge bases (KBs)using generated MRs. Previous works offen treat them in cascaded manner, that isfirst finishing the transformation from a natural lauguage question to their meaningrepresentation, then searching the answers using the generated meaningrepresentation. Unlike previous methods which treat them in order, we present atranslation-based approach to solve these two tasks in one unified framework whichcan handle them jointly.We translate questions to answers based on CYK parsing. Each CYK cellcorresponds to a question span, each question span as a meaningfull sub-questioncan be answered with KB entities. Answers as translations of the span covered byeach CYK cell are obtained by a question translation method, which first generatesformal triple queries as MRs for the span based on question patterns and relationexpressions, and then retrieves answers from a given KB based on triple queriesgenerated.A linear model is defined over derivations, and minimum error rate training isused to tune feature weights based on a set of question-answer pairs. Compared to aKB-QA system using a state-of-the-art semantic parser, our method achieves betterresults.