Research And Implementation Of Discourse Level Imagination Model

With natural language processing and text mining technology rapid development, people’s demand is also escalating, for example, the search engine, deep Q&A, this brought new challenges and opportunities, and also put forward higher requirements to the text processing technology. The main method of the challenge is the accurate semantic analysis technology. In the existing semantic analysis method, such as the semantic analysis of lexical level, phrase level an d sentence level are mainly aimed at the semantic of lexical level, phrase level and sentence level. This paper focuses on introducing text related external knowledge into the text to assist semantic analysis by making up the computer lacking of the text content relevant background information, in order to gain a better semantic analysis result of the text from the text level of semantic analysis.In this paper, according to introducing external information into text, we put forward three different models, referred to as imagination model, which are based on the iterative sequence of imagination model, based on the weight propagation of imagination model and merge discourse structure into imagination model. Imagination model are inspired by cognitive psychology, human reading habits, thinking habits and classical algorithm in information retrieval. We realizes that it is hard to sift through a large number of external information. Therefore, we have to select part of the external information to introduced into the text. Besides, in this paper, we also notice the application of external information and use it in text classification. The results of adding external information has achieved better results than the method without external information.The imagination model includes three parts: text representation, knowledge representation and model algorithm.In the text representation, we combine cognitive psychology research such as "Teachable-Language Comprehender"(TLC) in the design of the iterative sequences model, also combine the "Spreading-activation Model" in the design of weight propagation based imagination model.In addition this paper also pay attention to the internal structure of the text on introduction of the importance of external information. We join the discourse structure analysis such as the inter sentence relations recognition and Chinese mainstream CDTB corpus into the existing model, and use simple sentence relation recognition method to recognized discourse structure. After that we bring t he discourse structure information into the text representation scheme.In the knowledge representation, this paper uses triples as the representation method as with large semantic knowledge base. The triple representing knowledge concept of "Argument-Relation-Argument". By means of Latent Dirichlet Allocation(LDA) model, the word vector is quantized and the word vector is integrated into a vector of triple, and, at last, triples are represented as the continuous real-number vector.In the model algorithm, the model combines two classical models of the sequence model and the graph model which are the Page Rank algorithm and label propagation algorithm(LPA) and achieve better results by tuning the parameters of imagination model.To sum up, this paper is about using external information aided understanding perspective of discourse semantics of combining the classic algorithm in the cognitive psychology, text mining, and develop the imagination model. We also use the models in the text classification task and verify that external information of text semantic provides a feasible solution to the semantic understanding of the text.