Research On Implicit Sentiment Analysis Techniques Based On Attention Mechanism

Implicit sentiment analysis requires the intelligent system to predict the sentiment tendencies contained in the entire text without the obvious emotion words in the description document.In implicit emotion analysis,the key emotion words are missing and the emotion description of the target text is obscure requires the intelligent system to consider the overall information and context interaction characteristics of text from the global perspective,which is a difficult problem in the field of sentiment analysis.The existing analysis methods cannot extract the global emotion features of the input text while taking into account the effective extraction of local features,which leads to problems such as un-joint context modeling,difficulty in global feature extraction,and poor interpretability.This thesis focuses on the related problems and the challenges of text implicit sentiment analysis problems.We use two ways to divide the input text into different parts: by sentiment dictionary or natural break points in the conversation.Firstly,the partial feature is enhanced by local attention.Then,the global attention mechanism is used to model the overall important degree of features.This way try to solve the problem of sentiment analysis in the implicit text from different text granularity.The main contributions of this thesis are:(1)Implicit sentiment analysis based on the interactive attention mechanism.When the text lacks obvious emotion keywords,the feature representation of the input text poses serious difficulties in emotion analysis due to the lack of direct evidence features.This thesis presents an implicit emotion analysis model based on the context interaction attention mechanism.The model divides the input text into contextual features based on the segmentation points.By encoding the contextual features respectively,and using the context fusion matrix,the interactive attention mechanism is integrated to obtain the global emotional features of the whole text,and finally predicts the emotional tendency of the whole text.In this thesis,we further analyze the impact of the selection of segmentation points on the model performance.Finally,we use visualization techniques to explore the internal mechanism of the model’s predictive power.(2)Interactive emotion inference model for emotion recognition in conversation.Emotion recognition tasks in conversation can be widely used in empathy machines such as chat robots and intelligent agents.Efficient and accurate conversational emotion recognition models can improve the user experience.This thesis presents the interactive emotion inference model which uses interactive attention to fuse multiple rounds of conversation features,obtains the emotional feature representation of the entire conversation.During the high level feature extraction phase,the interactive emotion inference model extracts the long-distance emotional dependencies in the context by Transformer as an encoder.Then the predictive unit predicts the emotions contained in the conversation by using the above information.(3)Recurrent memory network model based on the interactive attention mechanismTo overcome the high storage cost and difficulty of training,we propose a recurrent memory network model based on the interactive attention mechanism.This model only requires one memory unit.The content of the memory unit filters the historical information according to the target task and inputs context text characteristics to update the memory unit.This design is more consistent with the general process of human emotional cognition,the whole structure is simple and effective.Experimental results show that the proposed model can effectively solve emotion recognition tasks in conversation.