基于结构-语义融合的评论文本情感分类研究

A study on sentiment classification of commentary text based on structural-semantic fusion

为了解决当前部分情感分析模型片面依赖文本语义特征、忽视句法结构特征的问题,提出了一种基于结构-语义融合的情感分类模型SF-TLSTM(sentiment-fusion-tree LSTM),即将句法结构信息与语义信息进行融合,全面提取文本特征。首先,将BERT(bidirectional encoder representation from transformers)模型引入TreeLSTM(tree-structured bidirectional LSTM)网络结构中;其次,利用SimCSE(simple contrastive learning of sentence embeddings)模型的自监督训练对BERT获得的向量表示进行数据增强;最后,通过节点编码的方式在构建的TreeLSTM网络上实现结构语义特征融合,并与基线模型进行多组对比分析。结果表明:在斯坦福大学发布的SST(stanford sentiment tree-bank)数据集上,SF-TLSTM模型相较于经典树形结构情感分类模型获得更高的准确率,其中在二分类任务中的准确率达到了96.79%。所提方法能够全面有效地提取评论文本的特征,增强文本的向量表示,对自然语言处理领域中的文本处理具有重要意义。

To solve the problem that some current sentiment analysis models unilaterally rely on textual semantic features and neglect syntactic structural features,a sentiment classification method based on structural-semantic fusion was proposed.The method fused syntactic structural information with semantic information to comprehensively extract text features.The BERT(Bidirectional Encoder Representation from Transformers)model was introduced into the TreeLSTM(Tree-structured bidirectional LSTM)network structure.Then the data enhancement of the vector representation obtained by BERT was performed by utilizing the SimCSE(Simple Contrastive Learning of Sentence Embeddings)model's self-supervised training.Finally the structural semantic feature fusion was realized on the constructed TreeLSTM network by means of node encoding,and was analyzed in a multi-group comparison with the baseline model.The experimental results show that on the SST(Stanford Sentiment Tree-bank)dataset released by Stanford University,the structural-semantic fusion-based sentiment classification method obtains higher accuracy compared to the classical tree-structured sentiment classification model,with an accuracy rate of 96.79%in the binary classification task.The proposed method can comprehensively and effectively extract the features of the comment text and enhance the vector representation of the text,which is important for text processing in the field of natural language processing.