Research On Opinion Mining For Customer Service Conversation Text

Customer service is one of main approaches for customers to give feedbacks,which contains a lot of valuable information for service quality monitoring,customer demand mining and product improvement.In the past,due to the poor speech recognition technology,the analysis of customer service depended on manual analyzing,which had problems of high cost and low coverage.With the advances of speech recognition technology,the data after speech recognition become more reliable and analysable.In this paper,three aspects are studied: named entity recognition,attribute word and sentiment word extraction and opinion collocation extraction in the customer service data after speech recognition.For named entity recognition,considering the problems of similar sound errors and few annotation data,this paper uses two fuzzy Pinyin features to adapt similar sound errors,and trains model with both partially annotated data and totally annotated data.Experimental results show that both fuzzy Pinyin features and additional partially annotated data can improve the performance of the NER model.Our proposed approach yields significant improvements over a CRF baseline and LSTM-CRF,achieving 0.79 F1 score on product name identification and 0.84 F1 score on organization name identification.For attribute words and opinion words extraction(opinion elements extraction),based on Double Propagation framework,this paper extracts attribute words and opinion words from both external corpus and customer service corpus.For external corpus,we design 11 kinds of dependent syntactic rules to extraction opinion elements.For customer service corpus,we combine the dependent syntactic rules and the rules based on window and part of speech to extract relations,filter relations having low frequency,and iteratively extract candidate opinion elements.Finally we filter low-frequency words and use word2 vec similarity to recall the valid low-frequency words to get the final attribute words and opinion words.Experimental results show that using external corpus and word2 vec similarity can both improve the performance of attribute words and opinion words extraction in customer service corpus.For opinion collocation extraction,this paper considers features about collocation rationality and collocation context in sentence.And we use Support Vector Machine to train opinion collocation relation classifier with these features.Our experiment results show that combine these two aspects of features can improve the performance of opinion collocation relation classifier.