Research On The Position Detection Method For Chinese Short Text Dialogue And Its Application

With the gradual maturity of Internet technology,there are a large number of user-generated content on the network,many of them are short text conversation.This make the stance detection for short text conversation become the research hotspot of opinion mining.Its purpose is to determining the position of the user from the text is in favor,against or either automatically.In this thesis,we proposed the stance detection task for short text conversation and proves the feasibility of the task from experiments result.Further,we integrate the stance information into response generation.The specific content is as follows:1.Construction the annotation corpus of stance detection in short text conversation:It's very important for our research to construct the corpus.We first collect over four million post-response pairs,and re-definition the target and stance for short text conversation.Then,we select the post-response pairs as to be labeled data,and propose the annotation instruction for the task.Finally,we obtain the large-scale,high-quality labeled corpus for the task of stance detection in short text conversation.2.Research on stance detection in short text conversation:We treat the task as a classification problem.Firstly,we use the max-entropy classifier based on discrete features to verify the feasibility our proposed task and report the baseline experiment of the task.Then,we use neural network based on dense features to improve the performance.In addition,we investigate the internal structures of stance information expressions and the correlation mechanism between stance and sentiment.Thus achieving the purpose of improve the theoretical basis of the stance detection tasks.3.Integrate stance information for response generation:In this thesis,we proposed a novel end-to-end neural model for response generation,by using a transition-based framework.First,we exploit various kinds of networks to encode source input sentences,and then integrate the stance information into neural model to build a stance-oriented response generation system.Experiment show that,the stance information an improve performance of response generation by decreasing the perplexity score.