{"title":"Graph Convolutional Networks with Dependency Parser towards Multiview Representation Learning for Sentiment Analysis","authors":"Minqiang Yang, Xinqi Liu, Chengsheng Mao, Bin Hu","doi":"10.1109/ICDMW58026.2022.00070","DOIUrl":null,"url":null,"abstract":"Sentiment analysis has become increasingly important in natural language processing (NLP). Recent efforts have been devoted to the graph convolutional network (GCN) due to its advantages in handling the complex information. However, the improvement of GCN in NLP is hindered because the pretrained word vectors do not fit well in various contexts and the traditional edge building methods are not suited well for the long and complex context. To address these problems, we propose the LSTM-GCN model to contextualize the pretrained word vectors and extract the sentiment representations from the complex texts. Particularly, LSTM-GCN captures the sentiment feature representations from multiple different perspectives including context and syntax. In addition to extracting contextual representation from pretrained word vectors, we utilize the dependency parser to analyse the dependency correlation between each word to extract the syntax representation. For each text, we build a graph with each word in the text as a node. Besides the edges between the neighboring words, we also connect the nodes with dependency correlation to capture syntax representations. Moreover, we introduce the message passing mechanism (MPM) which allows the nodes to update their representation by extract information from its neighbors. Also, to improve the message passing performance, we set the edges to be trainable and initialize the edge weights with the pointwise mutual information (PMI) method. The results of the experiments show that our LSTM-GCN model outperforms several state-of-the-art models. And extensive experiments validate the rationality and effectiveness of our model.","PeriodicalId":146687,"journal":{"name":"2022 IEEE International Conference on Data Mining Workshops (ICDMW)","volume":null,"pages":null},"PeriodicalIF":0.0000,"publicationDate":"2022-11-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"2022 IEEE International Conference on Data Mining Workshops (ICDMW)","FirstCategoryId":"1085","ListUrlMain":"https://doi.org/10.1109/ICDMW58026.2022.00070","RegionNum":0,"RegionCategory":null,"ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"","JCRName":"","Score":null,"Total":0}
引用次数: 0
Abstract
Sentiment analysis has become increasingly important in natural language processing (NLP). Recent efforts have been devoted to the graph convolutional network (GCN) due to its advantages in handling the complex information. However, the improvement of GCN in NLP is hindered because the pretrained word vectors do not fit well in various contexts and the traditional edge building methods are not suited well for the long and complex context. To address these problems, we propose the LSTM-GCN model to contextualize the pretrained word vectors and extract the sentiment representations from the complex texts. Particularly, LSTM-GCN captures the sentiment feature representations from multiple different perspectives including context and syntax. In addition to extracting contextual representation from pretrained word vectors, we utilize the dependency parser to analyse the dependency correlation between each word to extract the syntax representation. For each text, we build a graph with each word in the text as a node. Besides the edges between the neighboring words, we also connect the nodes with dependency correlation to capture syntax representations. Moreover, we introduce the message passing mechanism (MPM) which allows the nodes to update their representation by extract information from its neighbors. Also, to improve the message passing performance, we set the edges to be trainable and initialize the edge weights with the pointwise mutual information (PMI) method. The results of the experiments show that our LSTM-GCN model outperforms several state-of-the-art models. And extensive experiments validate the rationality and effectiveness of our model.