摘要： 方面级别文本情感分析旨在分析文本中不同方面所对应的情感趋向。传统基于神经网络的深度学习模型在文本情感分析的过程中，大多直接使用注意力机制而忽略了句法关系的重要性，即不能充分利用方面节点的上下文语义信息，导致情感趋向预测效果不佳。针对该问题，设计一种融合句法信息的图注意力神经网络模型，并将其应用于文本情感分析任务。利用预训练模型BERT进行词嵌入得到初始词向量，将初始词向量输入双向门控循环神经网络以增强特征向量与上下文信息之间的融合，防止重要上下文语义信息丢失。通过融合句法信息的图注意力网络加强不同方面节点之间的交互，从而提升模型的特征学习能力。在SemEval-2014数据集上的实验结果表明，该模型能够充分利用句法信息进行情感分析，其准确率较LSTM及其变种模型至少提升3%，对Restaurant评论进行情感分类预测时准确率高达83.3%。

关键词: 情感分析, 门控循环神经网络, 图卷积神经网络, 句法依赖, 注意力机制

Abstract: Aspect-level text sentiment analysis is oriented to emotional trends corresponding to different aspects of a text.Most of the traditional deep learning models based on neural networks use attention mechanism directly, and ignore the importance of syntactic relations, which makes them underutilize contextual semantic information of aspect nodes and hurts the performance in emotional trend prediction.To address the problem, a graph attention neural network model integrating syntactic information is proposed and applied to textual sentiment analysis tasks.The model employs the pre-trained model BERT for word embedding to obtain the initial word vectors, and the initial word vectors are input into a two-way gate control loop between neural network to enhance the fusion of eigenvector and context information, so the loss of important contextual semantic information can be avoided.Then the graph attention network integrating syntactic information is used to strengthen the interactions between different aspect node, and thus strengthen the ability of the model to learn features.Experimental results on the SemEval-2014 dataset show that the proposed model can utilize syntactic information for sentiment analysis.Its accuracy is at least 3% higher than that of LSTM and LSTM variations, and reaches 83.3% in sentiment classification tasks for Restaurant reviews.

Key words: emotion analysis, gate Recurrent Neural Network(RNN), Graph Convolutional Network(GCN), syntactic dependence, attention mechanism

中图分类号: 

• TP391