A structure-enhanced graph convolutional network for sentiment analysis

Findings of the Association for Computational Linguistics: EMNLP 2020, pages 586–595 November 16 - 20, 2020. c⃝2020 Association for Computational Linguistics 586 Sentiment Analysis with Weighted Graph Convolutional Networks Fanyu Meng, Junlan Feng, Danping Yin, Si Chen, Min Hu China Mobile Research Institute, Beijing, China {mengfanyu,fengjunlan,yindanping,chensiyjy,humin}@chinamobile.com Abstract Syntactic information is essential for both sen- timent analysis(SA) and aspect-based senti- ment analysis(ABSA). Previous work has al- ready achieved great progress utilizing Graph Convolutional Network(GCN) over depen- dency tree of a sentence. However, these mod- els do not fully exploit the syntactic informa- tion obtained from dependency parsing such as the diversified types of dependency rela- tions. The message passing process of GCN should be distinguished based on these syntac- tic information. To tackle this problem, we de- sign a novel weighted graph convolutional net- work(WGCN) which can exploit rich syntactic information based on the feature combination. Furthermore, we utilize BERT instead of Bi- LSTM to generate contextualized representa- tions as inputs for GCN and present an align- ment method to keep word-level dependencies consistent with wordpiece unit of BERT. With our proposal, we are able to improve the state- of-the-art on four ABSA tasks out of six and two SA tasks out of three. 1 Introduction Sentiment analysis(SA), also known as opinion mining, is the task of determining the polarity of a piece of text. Commonly the classification is whether the text is expressing a negative or positive attitude towards a topic or a product. Fine-grained sentiment analysis involves more than two senti- ment classes (very negative, negative, neutral, pos- itive and very positive). Aspect-based sentiment analysis(ABSA) is one step further by assigning sentiment polarities to specific aspects of an in- volved entity or a topic. For example, comment on a restaurant saying “The restaurant was expensive, but the menu was great” has positive and negative attitudes for two aspects food and price. Much progress has been made recently to ad- vance the state-of-the-art on shared SA and ABSA tasks. Contributions mainly come from two re- search directions. One is to take advantage of the pre-trained lan- guage models such as ELMo(Peters et al., 2018), BERT(Devlin et al., 2018) and XLNet(Yang et al., 2019a), which are typically employed to extract contextual features of a piece of text for the final classifer. These models effectively alleviate the heavy effort of feature engineering of earlier work on SA and ABSA. Further inventions have been proposed to better fine-tune these models. For in- stance, a recent work (Sun et al., 2019a) converts ABSA to a sentence pair classification task, where an auxiliary sentence is generated. It then fine- tunes the pre-trained model from BERT for this new task. Promising experimental results are ob- served. Second line of research is to exploit the syntactic structures of subjective sentences with a belief that interactions between words need to be considered in sentiment analysis, which however is not suffi- ciently captured by even the latest attention-based models. (Zhang et al., 2019) quotes a concrete example “Its size …

Content truncated to 500 words