Research On Aspect-based Sentiment Analysis Of Complaint Dialogue Of Telecommunication Carrier

In the field of natural language processing,Aspect-based Sentiment Analysis is an important fine-grained research task,which is widely used in understanding public opinion,market research,brand reputation Analysis,customer experience identification,evaluation of social media influence and other fields.It aims to analyze the polarity of emotions corresponding to multiple aspects of a text or sentence.The main difficulty of this task lies in the lack of a large number of labeled data to aspect-level sentiment analysis,and the existing actual data are noisy,have not strong enough regularity,aspect words inclusion and overlapping.Compared with traditional methods,deep neural network based method can extract rich text features and improve the task completion effect.However,many existing methods fail to capture the effective contextual information accurately,and can not make full use of the association information between aspect words and their adjacent contexts.However,in a large number of previous studies,aspect-level sentiment analysis is divided into two steps: extracting aspect-level words first and then judging the emotional polarity of a particular aspect.Ignoring the relationship between the two subtasks will lead to error transmission,making the model further judge the emotional polarity of the aspect words that are already misidentified.Therefore,aspect word extraction and specific aspect affective polarity judgment tasks are mutually promoting.How to simultaneously extract aspect word and judge its affective polarity using the correlation information between the two sub-tasks is extremely challenging.Considering that the method of multi-task joint learning can optimize the model parameters in the back propagation of the loss function of different subtasks,the subtasks can influence each other.The local context focus mechanism can capture the context information of aspect words well.In order to adapt to the situation where multiple words or aspect words are included and overlapped in a sentence,and to make the model adapt to the strong domain of carrier complaint dialogue more fully,this paper proposes an aspect level sentiment analysis model based on the span annotation BERT of carrier complaint text pre-training,which is specifically developed from the following two parts:(1)Based on the pre-training language model BERT,this paper first proposes to use unlabeled operators’ complaints data to conduct the pre-training of domain information transfer based on the masked language model.The probe task is used to verify that the domain information transfer of operators’ complaints can effectively improve the effect of aspect word extraction task.In this paper,we explore the implicit representation of domain,aspect and non-aspect words generated by the pre-trained language model with domain transfer by using probe test,and select some samples to visually display the distribution of the implicit representation under the two kinds of classification by dimensionality reduction method.Under these two distinctions,the implicit representations generated by BERT after domain migration have relatively obvious subspaces.In addition,this paper uses the span based annotation method to extract aspect words on this basis.By comparing with several more effective benchmark methods,the model in this paper has obtained more accurate recognition of aspect words in complex texts.(2)In order to capture the association information between aspect words and their emotional polarity,this paper proposes a BERTp-LCF-ATSEPC model for end-to-end aspect sentiment analysis of carrier complaint texts.By simultaneously extracting aspect words and judging the corresponding affective polarity,the multi-task learning model can make better use of the correlation information between aspect word context and affective polarity.The local context focus mechanism is applied to make the fine-grained interaction between aspect context representation and affective polarity,and the judgment of specific aspect emotional polarity is obtained.In addition,this paper adopts span based annotation to deal with the problems of inclusion and overlap of aspect words and multi-faceted words.Experiments on complaint dialogue text data show that BERTp-LCF-ATSEPC model can capture the relationship between aspect words and emotional polarity more fully,and further improve the performance of aspect level sentiment analysis in carrier complaint dialogue compared with the current advanced end-to-end methods.