Research And Application On People's Opinion And Activity Extraction System

People's opinion and activity extraction aims to extract viewpoints and activity information of target person from unstructured data. Due to the variety of activities, this thesis mainly concerned about the related event as a typical activity of target person. This system has important application value for tasks such as automatic compilation of characters' chronology and expert finding. The detail to realize this system is as follows.First, this thesis designs a Chinese temporal expression extraction method based on words and part of speech to extract time information.Experimental results show that the method has good performance of time extraction.Second, this thesis proposes a two-phase method for people's opinion extraction. The first phase is recognition of people's opinion sentence which is the key stage of the method, and the second phase is people's opinion extraction. The experiments investigate differences between classifiers such as SVM, Logistic regression and features such as trigger,relative position and dependency relation for recognition of people's opinion sentence. Some useful conclusions are given through experiments.Then the two-phase method mentioned above is applied to people's event extraction task, that is, event recognition is carried out before event extraction. On the basis of the characteristic of peoples' events, a one-stage method is put forward. The one-stage method has no need of recognition of people's event sentence, and directly extracts events of target person. In this paper, long-short term memory model is applied to two-phase and one-phase method of people's event extraction. Experiments show that the performance of one-stage method is slightly better than two-stage method.Finally, this thesis achieves a complete people's opinion and activity extraction system cascading opinion extraction, event extraction and temporal Chinese expression extraction methods referred before. The system extracts opinion and related activities from the network text according to the names of the characters inquired by the users and returns structured results to the users and stores them in database.