Research And Application On Emotion Classification Of Customer Service Speech Based On Bimodal

A technology company has saved a large amount of customer service recording data,but it needs to change the traditional manual quality inspection method with poor effect and low efficiency to improve service quality.In view of the current advantage of deep learning technology in dealing with emotional classification issues with high accuracy and efficiency,and in combination with the characteristics of the company’s voice data,in the project,monitoring and analysis of customer service voice based on the bimodal model were carried out to effectively improve the company’s customer service level.In this thesis,first of all,starting with the bimodal characteristics of recorded files,a multimodal alignment method based on an improved attention mechanism was adopted to correlate the two modal features obtained after converting customer service recorded files into text and speech,and form the final emotional expression;Then,in response to the emotional value bias caused by the fact that the basic emotional dictionary does not have the corporate precursor chemicals theme,the thematic emotional dictionary was constructed to optimize the text data;After using a topic emotion dictionary to calculate emotion values to enhance the emotional orientation of the text,the optimization of the emotion classification algorithm was completed,and the effect was verified through experiments;Finally,relying on the above algorithm,the project requirements were completed.The main work of this thesis was as follows:(1)In order to solve the problems existing in traditional single modal emotion analysis methods,the bimodal emotion analysis method was adopted to convert customer service recordings into two modals of data and extract their features.On the one hand,the Word2 Vec algorithm was used to extract text features;On the other hand,Mel Frequency Cepstral Coefficients were used to represent speech features.(2)Aligning the features of both text and speech modals through attention mechanisms.In this thesis,the multimodal alignment method based on self-attention was adopted to improve the self-attention mechanism,calculate the alignment weights of the two modals,capture the implicit associations between different modals,enhance their complementarity,and reduce redundancy,thereby obtaining more specific feature representations for more accurate classification in subsequent classification work.(3)Through the algorithm improvement of the Semantic Orientation Pointwise Mutual Information,the expansion of the emotion dictionary was completed to obtain the emotion dictionary based on the company’s precursor chemicals theme.Then,the topic emotion dictionary based calculation method was used to obtain the emotional scores of each text data,thereby enhancing the presentation effect of text features,and ultimately using new text features as input to the algorithm to improve algorithm performance.(4)Design and implementation of quality inspection module in company information system.Relying on the algorithm in this thesis,the design and code implementation of the quality inspection function and related functions derived from the quality inspection function in the system were completed.After testing,verification,and online operation,it can better meet the company’s needs for customer service quality inspection applications.