摘要： 针对现有文本情感分析方法存在的无法高效捕捉相关文本情感特征从而造成情感分析效果不佳的问题，提出一种融合双层多头自注意力与卷积神经网络（CNN）的回归模型DLMA-CNN。采用多头自注意力机制学习序列内部的词依赖关系，从而捕获序列的内部结构。重利用浅层特征并与多头自注意力特征进行融合，结合深度学习中的CNN进一步优化文本情感极性分析效果。在基准数据集SemEval-2017 Task 5上进行实验，结果表明，与传统机器学习算法CNN、ELSTM、Att-BLSTM等相比，该模型取得了较好的情感极性分析效果，且运行效率较高。

关键词: 金融文本, 情感极性分析, 多头自注意力机制, 特征融合, 深度学习

Abstract: Existing text sentiment analysis methods fail to efficiently capture the emotional features of the texts,which weakens the performance of sentiment analysis.To solve the problem,this paper proposes a regression model,DLMA-CNN,which uses the multi-head self-attention mechanism to learn the word dependency relationships within the sequence and capture the internal structure of the sequence.The model reuses shallow features and fuses them with multi-head self-attention features,using Convolution Neural Network(CNN) in deep learning to achieve better performance of sentiment polarity analysis in texts.Experiments on the benchmark dataset SemEval-2017 Task 5 show that the proposed model outperforms CNN,ELSTM,Att-BLSTM and other traditional machine learning algorithms in terms of sentiment polarity analysis,and can run more efficiently.

Key words: financical text, sentiment polarity analysis, multi-head self-attention mechanism, feature fusion, deep learning

中图分类号: 

• TP18