融合RoBERTa和注意力机制的隐喻方面级情感分析

Metaphorical Aspect Sentiment Analysis Based on RoBERTa and Attention Mechanism

针对目前大多数隐喻情感分析方法存在对方面情感注意力引入不足的问题,提出一种用于隐喻方面级情感分类的模型.模型首先通过RoBERTa对具有方面情感信息的文本进行编码,将编码后的方面信息和多层情感注意力信息融合,形成多层方面注意力表征向量.将该表征向量与隐喻句的关联结果作为文本原始特征,利用注意力机制和方面信息对其解码,然后通过卷积网络计算隐喻句与方面词的关联度.将池化层输出结果和卷积计算结果合并,最后计算隐喻句不同方面词的情感极性的概率,完成隐喻情感分析.实验结果表明该模型对3种情感极性的平均判断准确率分别达到了83.26%,81.69%和56.68%,与基线实验相比均有所提升.

Aiming at the problem that most of the current metaphor sentiment analysis methods lack the introduction of aspect emotional attention,a model for metaphor aspect-level sentiment classification is proposed.The model first encodes the text with aspect emotional information through RoBERTa,and fuses the encoded aspect information and multi-level emotional attention information to form a multi-level aspect attention representation vector.The correlation between the representation vector and the metaphorical sentence is used as the original feature of the text,and the attention mechanism and aspect information are used to decode it.Then the correlation between the metaphorical sentence and aspect words is calculated by convolution network.The pool layer output results and convolution calculation results are combined.Finally,the probability of emotional polarity of words in different aspects of metaphorical sentences is calculated to complete metaphorical sentiment analysis.The experimental results show that the average judgment accuracy of the model for the three emotional polarity reaches 83.26%,81.69%and 56.68%,respectively,which is improved compared with the baseline experiment.