Research On Fine-Grained Sentiment Analysis Method For Supporting Automobile Service Quality Evaluation

The integration of service and manufacturing is the developing trend of the current automotive industry.As a consequence,in order to obtain new growth points and profit sources for the automotive industry,it is a certain way to provide more profound and more fine-grained services for the huge car ownership market.Meanwhile,obtaining the customer's comprehensive sentimental experience of each service is the premise and key to the improvement of service quality.The customer's comprehensive sentimental experience includes coarse-grained sentiment,which only shows white or black opinions,and should also include a specific and clear fine-grained sentiment.With the increasing competition in the market and the importance of satisfying the individualized needs of customers,it is becoming more and more obvious that the defects of the past coarsegrained sentimental evaluation methods are challenging to capture the specific needs of customers.As a result,the customer's access to the clear and specific fine-grained sentimental experience of service quality has become an urgent need of the current automotive service enterprises,and an essential basis for these enterprises to provide targeted services.To this end,this paper focuses on the customer service evaluation data in the past 10 years,which covers pre-sale,sale and after-sale service of automobile service enterprises,in the third-party "ASP/SaaS-based Manufacturing Industry Value Chain Collaboration Platform"(hereinafter referred to as the platform)in the national key R&D program "Distributed Resource Giant System and Resource Synergy Theory"(Projection No.: 2017YFB1400301).In these evaluation data,the unstructured text data of the subjective sentiment is used to focus on the fine-grained sentiment evaluation method of automobile service quality.The thesis mainly includes:(1)The state of service quality evaluation of service enterprises in the platform is firstly analyzed.According to the incompleteness of user opinions and the coarse granularity of analysis conclusions caused by the fragmentation of data on the platform,and on the basis of the research on the sentiment analysis methods of existing textoriented evaluation data,a fine-grained sentiment analysis scheme is proposed,which consists of text preprocessing,two-stage sentiment elements extraction and sentiment elements de-weighting.(2)According to the problem of randomness,mixed data,non-standardization in the user evaluation text data,and the continuous uninterrupted word and word arrangement description of Chinese text,the evaluation text is preprocessed by invalid text filtering technology,Chinese word segmentation and word2 vec model respectively to provide data quality assurance and formal support for text data in subsequent sentiment extraction.(3)According to the lack of labelling knowledge of user evaluation and the characteristics of existing sentiment element extraction methods,a two-stage sentiment elements extraction method is proposed,which is based on syntactic rule matching and deep semantics and literal distance.An algorithm based on dependency grammar and subject-predicate relationship is proposed in the first stage,and it is performed on the unlabeled text dataset to obtain the preliminary opinion expressions and the aspect of the opinion.Moreover,then these opinion expressions and the aspect of the opinion are formed into tagged knowledge datasets.Also,an extraction and matching algorithm based on depth semantics and literal distance is proposed in the second stage.Meanwhile,the tagged knowledge dataset formed in the first stage is also used when training.The extraction algorithm employs deep recurrent neural network combined with semantic features to further extract more sentiment expressions accurately.However,the extraction result contains multiple opinion expressions and the aspect of the opinions to be matched.As a consequence,the matching algorithm based on the minimum literal separation distance and the opinion expression is proposed to perform pairwise matching.Experiments verify the effectiveness of the method.(4)Aiming at the sentimental redundancy caused by the synonym repetition in some preliminary conclusions of the two-stage algorithm,the semantic de-weighting algorithm based on regularized AutoEncoder and K-means clustering is proposed to obtain more accurate and precise fine-grained sentiment elements,and experiments verify the method.Based on the above method,the fine-grained sentiment extraction of the evaluation text in this paper is completed.