Research On The Extraction Of Opinion Targets For Chinese Short Text Of Comment

Emotional tendency analysis has widely application in practice. For short text of comment, accurately tagging opinion word and its opinion target plays an important role in determining emotional tendency. Because of the absence of morphological changes and flexible relation modifications in Chinese, tagging opinion target has still not been solved perfectly. With the development of dependency parsing techniques for Chinese sentence, short text of comment now could be correctly parsed into dependency relation tree, and so specific natural language patterns reflecting relationships between opinion target and its features could be mined, which could be applied to improve the performance of opinion target tagging method.The main works of this dissertation are as follows:(1) On the basis of the observation to the existing opinion target tagging algorithms, based on frequent tree patterns mined from dependency relation tree bank, a novel opinion target tagging method for Chinese short text of comment has been proposed. This method includes three main steps, which are initial opinion target tagging step based on frequent tree patterns, ordered frequent tree patterns and rule set training step based on error-driven TBL framework, and opinion target tagging step using ordered frequent tree patterns and rule set. Differently from traditional methods, FTTBL could automatically extract features related to opinion target to compose frequent tree patterns. And what is more important is that FTTBL employs error-driven TBL framework to successfully solve the choice and fuse of frequent tree patterns and rules. Experimental results show the effectiveness of this method.(2) For time-consuming problem in ordered frequent tree patterns and rule set training step in FTTBL, two schemes have been designed. First is the reduction of initial frequent tree patterns and rule set. Redundancy in initial frequent tree patterns and rule set has been studied in detail and redundancy eliminating scheme which could assure the tagging accurancy has been designed. For the training step of ordered frequent tree patterns and rule set, the most time-consuming step in FTTBL, the studies show that the main cause is the duplicate computation between adjacent two iterations. So the characters of the duplicate computation in the training steps have been studied in detail, and then the scheme of using some additional storage to reduce duplicate computation has been designed to finally reduce the computation time. Experiments have been conducted to show the effectiveness of the schemes.