Reasearch And Application Of Sentence-Level Dialogue Sentiment Analysis

Dialogue text sentiment analysis is a research direction of natural language processing,which calculates the sentiment characteristics contained in texts to obtain its sentiment tendency.Different from traditional sentiment computing tasks,dialogue text contains information interaction.On the basis of analyzing the user's emotional tendency,it is necessary to study the changes and effects of emotions between users in the dialogue,which exists many technical difficulties.This topic builds a sentiment classification model,which based on dialogue texts,to achieve the sentiment classification task of interactive dialogue texts.The main work of this paper are as follows:(1)A character-level OOV solution is proposed,which solves the impact of unregistered words on the construction of corpus vocabulary during the text preprocessing stage.This topic deals with the input corpus from the perspective of corpus analysis and text vectorization,and constructs auxiliary vocabularies such as slang replacement vocabulary,expression replacement table,corpus vocabulary,etc.Besides,we use character-level OOV solution to fine-grain the text and preserve the original information of the corpus.The word2vec model is used for domain training on both word and expression vectors,and finally get a vectorized text representation adapted to the subject.(2)A Multi-Dimensional Feature Extraction Model(MDFE model)is proposed to realize the emotional feature extraction of dialogue texts in both temporal and spatial dimensions.The MDFE model introduces the attention mechanism,which simulates the dialogue text generation process in the time dimension and captures the emotional changes in the user's communication process.The convolution operation is used in the spatial dimension to extract the structural characteristics of the text.Besides,the attention scoring mechanism is used to refer to the time and space dimensions,achieving multi-dimensional emotion feature fusion.(3)A Hierarchical Multi-Class Classification Model(HMCC model)is proposed to complete the emotion category prediction of dialogue text.The HMCC model builds a hierarchical classification system of "two classifications->multiple classifications" to solve the problem of model overfitting,which caused by uneven data distribution.We innovatively propose a Fusion Multi-Class Loss Function(FMC loss),which improves the loss calculation method of the multi-class model,and thereby promotes the classification judgment result and calculation efficiency of the classification model.(4)The design and implementation of dialogue text sentiment classification system.The system includes three modules:text preprocessing,emotion feature extraction and emotion category judgment,which realizes the function of emotion classification of dialogue texts.A visual system interface is built to realize the information exchange between the system and users,which intuitively demonstrated the functions of the system.