Research On Recommendation Methodology For Online Negative Reviews Handling

Online negative reviews posted publicly will spread quickly across social media platforms,damaging the image and reputation of enterprises,increasing the perceived risk of potential consumers,lowering the trust and purchase intention of potential consumers for the product or service,and ultimately bringing down the sales of the enterprise’s products and services by a large margin,or even changing the direction of development of the enterprise.If not handled in a timely and effective manner,online negative reviews will pose a huge threat to the development of the enterprise.Traditional manual methods fail to meet the requirements of online negative review handling due to poor efficiency,high cost,slow speed,and large resource consumption.This paper,focusing on the issue of fast and efficient handling of online negative reviews by effectively using the internal and external resources of the enterprise,proposed the value co-creation model for handling online negative reviews featuring multi-party participation and the corresponding value co-creation framework,which handle online negative reviews with two recommendation methods,namely expert recommendation and case recommendation.The main research contents and innovations are summarized as follows:(1)In terms of online negative review recognition,a two-stage model based on sentimental elements is built.In stage 1,the multi-layer Bi-LSTM-CRF model is used to mark the sentimental elements in online reviews.The input layer adopts word vectors,part-of-speech vectors and dependency vectors to express word features more specifically.In the multi-layer Bi-LSTM-CRF model,Bi-LSTM memorizes the current word and the semantic information of the context in both long-and short-term from two different directions,and uses the contextual semantic information to predict the sentimental element mark score of the current word.In the model,the CRF layer achieves the global optimum in terms of the output of the sentimental element marks for the total series.The experimental results indicate that the model can well mark sentimental elements.Stage 2 employs the domain-specific sentiment lexicon and the discriminative process to identify online negative reviews and build an improved domain-specific sentiment lexicon with alleviated polarity immobilization.The experiment verifies that the two-stage model can effectively and automatically identify online negative reviews.(2)In terms of the prioritization for handling online negative reviews,a model is built for the prediction of the usefulness of online negative reviews.The model predictors are composed of features of the reviews such as the number of words in the review,the sum of the useful values of words,and the number of product features.Among them,the useful value of a word in a specific domain is calculated by its information gain and multi-value bias.The loss function of the prediction model is also established.The model parameters are obtained using the gradient descent method based on massive online review data.According to the experimental results,the model can better predict the usefulness of online negative reviews.(3)Targeting the problem of identification of experts on online platforms and specific review-based recommendations of suitable experts for handling online negative reviews,the expert identification model and the expert recommendation model are built for handling online negative reviews.Different from previous models,the identification model takes not only the product knowledge of an expert but also the expert’s loyalty to the brand products,his/her community influence,community support,emotional appeal,time and energy into consideration when handling online negative reviews.Moreover,it obtains real data from the Xiaomi community and applies the artificial neural network to verify the validity of the expert identification model.In terms of expert recommendation,a model integrating knowledge matching,emotional solace matching,and time and energy matching is built.Based on the product knowledge ontology established,as well as the domain-specific knowledge keywords and feature vectors of the product,a knowledge model of experts’ knowledge in different domains is built.With the domain knowledge and domain knowledge correlation matrix,the explicit and tacit knowledge demands of online negative review handling are calculated.Emotional solace matching measures an expert’s emotional solace against the intensity of negative sentiments in an online negative review;time and energy matching is used to measure whether an expert has enough time and energy to handle negative online reviews.The effectiveness of the recommendation model is verified through experiments.(4)As for auto-reply to online negative reviews,a sentence similarity-based case recommendation method is proposed.According to the characteristics of online negative reviews,a sentence-based case library of online negative reviews is established,which adopts sentence similarity to retrieve the corresponding solution.To enhance the retrieval accuracy,a sentence similarity calculation integrating multi-features is proposed.The method analyzes a sentence from the perspectives of words,semantics,and semantic dependency,while taking common words,modifiers,keywords,semantics,semantic dependency association pairs and other features of the sentence into consideration.Experiments verify that the calculation method shows higher accuracy than other calculation methods of Chinese sentence similarity,guaranteeing the effect of case recommendation for handling online negative reviews.