Context-aware Heterogeneous Graph Convolutional Network For Sentiment Analysis

With the increasing demand for efficient and automated processing of massive text data,natural language processing tasks such as text classification and text sentiment analysis have emerged.This paper uses a deep learning model to deal with the problem of implicit sentiment analysis of text.By modifying and combining classical neural network models and using migration learning,this paper improves the effectiveness of deep models on implicit sentiment analysis tasks.First,to address the problem of lack of sentiment vocabulary in the target text and difficulty in discriminating sentiment polarity in implicit sentiment analysis,this paper proposes a Context-aware Semantic(CAS)model based on heterogeneous graphs for perceptual context.The CAS model is divided into four parts: 1)Extracting the features of the target text sentence and its context using a bidirectional GRU network,respectively.2)Fusing the syntactic dependency tree information,TF-IDF values of inter-sentence words,and inter-sentence order information of the target text sentence into the same graph employing heterogeneous graphs,and constructing a graph with the syntactic dependency tree information of the target text sentence.3)Use the graph convolutional network to extract features based on the graphs of contextual text and target sentence text in both directions.4)The features extracted from the target text sentence and the contextual text are fused using a hierarchical attention mechanism and input to the forward network,and the sentiment polarity distribution is output by a Softmax classifier.Compared with the compared models in this paper that do not consider contextual text information and do not use heterogeneous graphs,the experimental results of the CAS model on the dataset SMPECISA 2019 have an F1 value of 0.8256.Compared with the compared models,the CAS model effectively improves the prediction of sentiment polarity.Secondly,to solve the problem of the lack of implicit sentiment text dataset,this paper proposes a transfer learning method to transfer knowledge from the explicit text sentiment model to the implicit sentiment model,called the multi-grain size discriminator method.The multi-granularity discriminator method achieves inter-domain knowledge migration by fuzzing the discriminative ability of the implicit sentiment model for data between different domains employing gradient adversarial,and at the same time achieves knowledge migration between data of the same sentiment tendency by aligning the data labels between the same sentiment tendency.Meanwhile,to stabilize the training instability caused by domain adversarial in the multi-granularity discriminator,this paper introduces a mutual learning mechanism between implicit and explicit classification models to further realize inter-domain knowledge migration.Through comparison experiments from four different domain datasets transfer to the dataset SMP-ECISA 2019,we verify the comparative effectiveness of the proposed method compared with existing migration learning methods and the improvement of model prediction for implicit sentiment classification tasks.Finally,this paper implements a text sentiment analysis system using CAS model based on Flask Web development framework in Python language,deep learning framework Pytorch.