Short Text Semantic Extension And Sentiment Polarity Analysis Research

With the development of Internet and the popularization of smart phone,form of information and data becomes much more fragmented than before.Facing with large amount of short text data,how to understand its semantics in a more efficient way and improve the performance of emotional polarity analysis considering its data sparseness is an urgent problem to be solved.Text emotional polarity is a process of analysis,processing,induction and inference of subjective texts with emotional colors.This thesis chooses emotional polarity of short text as the driving problem,and takes effort to expand short text semantics in three aspects: unknown word recognition,text topic mining and feature extraction based on deep learning,which improves the performace of emotional polarity analysis of short text.The main work of this thesis is descriped as follows:1.An unknown word identification algorithm called NC-value for short is proposed which is a mixture of rule-based method and statistics-based method.The algorithm improves the recognition effect and segmentation quality of unknown words in short text preprocessing.Since traditional C/NC-value algorithm can not accurately identify the boundaries of unknown words and low-frequency words which is common in short text,the proposed method took the characteristics of unknown words in short text into consideration and combined the benefit of mutual information and branch entropy which helps to improve the performance of semantic mining and emotional polarity analysis.2.A short text enrich model called Conceptual and Semantic Enrichment with Topic Model(CSET)is proposed based on topic model which combines topic model Biterm Topic Model(BTM)with probabilistic knowledge graph Probase.CSET mines terms and concepts in short texts through Probase,and predicts semantic relationship between terms and concepts through topic model.CSET improves semantic analysis capacity of short text representation model.3.On the basis of the above work,this thesis proposes a short text emotional polarity analysis model based on semantic extension called SESA.Deep learning is used in the model to mine short text features which combines the traditional convolutional neural network model with the bi-directional long short term memory artificial model and attention-based pooling.The proposed model utilizes the advantages of the two models are to do semantic learning and deep feature mining for Short Texts and improves the effect of emotional analysis of short text.The above algorithms and models are validated by experiments in corresponding chapters.Experimental results prove the validity of these algorithms and models.