基于主动学习的语义缺失问句补全-计算机科学与技术专业论文.docxVIP

哈尔滨工业大学工学硕士学位论文 哈尔滨工业大学工学硕士学位论文 II II Abstract Interactive question and answering makes up many defects traditional search engine and normal question and answering have. Compared to the former, it could provide more relavent and more precise information. Compared to the latter, it has the better ability of interaction and natural language understanding so the high quality result is returned. Interactive question and answering needs to process contextual information so the ability of understanding is enhan ced. In context, the research about lack-of-semantic questions is necessary to do as one of core problems. There exists many lack-of-semantic questions in interactive question and answering system and also in the real Chinese context. According to statistics of the corpus in real environment by wizard-of-oz, about 45 percent of questions are lack-of-semantic questions. Without recovering missing information, the system could not get the user intention. Thus, the recovery of ellipsis questions is a crucial s tep in process of interactive question and answering. The questions are divided into four categories which are semantic integrity category, noun ellipsis category, reference category and other category by machine learning methods. Firstly, use the morphology and syntax information to extract feature and then train the classifiers and compare them. At last, the ID3 method is chosen and the precision reaches 80.1%. After classification, the recovery-completion model is put forward to define the searching space of antecedents. To deal with different categories, use corresponding methods to recover questions. Use the syntax and semantics information to extract features, train the classifier and determine the antecedent. Noun ellipsis category and reference category are processed by SVM classifiers. Experiment is on the corpus of manual collection and TREC. Then the result of recovery of questions are evaluated. By the experiment on the corpus in TREC, the