基于多任务和自注意力机制的文本微情感分析研究

Research on Text Micro Emotion Analysis Based on Multitasking and Self Attention Mechanism

基于多任务和自注意力机制的文本微情感分析研究,旨在挖掘句子中涉及的(4分类、7分类、28分类)的情感,特别是挖掘句子在28分类中的微情感。现有的文献大多仍集中于粗粒度(4分类)的情感研究。而对于7分类、尤其是28分类的微情感研究极少。为了解决上述问题,并且弥补国内对于微情感(28分类)方面研究的空白,论文提出了基于多任务和自注意力机制的微情感分析模型,即通过对4分类、7分类和28分类的三个情感分析任务同时处理,以其中某一个任务为主任务(给予0.9的权重),另外两个任务作为辅助任务(分别给予0.05的权重),并共享三个任务的网络权重,以达到提升主任务模型的目标。通过实验证明,对多个相关联的不同粒度的任务进行编码层的网络共享,能够提高模型对于句子的文本特征提取能力,从而提高模型对于句子的情感识别精度,尤其是在细粒度(微)情感分类任务中。

The research on text micro emotion analysis based on multitasking and self attention mechanism aims to mine the emotions involved in sentences(4 classification,7 classification and 28 classification),especially the micro emotions of sentences in 28 classification.Most of the existing literature still focuses on coarse-grained(4 classification)emotion research.However,there is little research on micro emotion of 7 categories,especially 28 categories.In order to solve the above problems and make up for the gap in the domestic research on micro emotion(28 classification),this paper proposes a micro emotion analysis model based on multi task and self attention mechanism,that is,through the simultaneous processing of three emotion analysis tasks of 4 classifi-cation,7 classification and 28 classification,one of them is the main task(with a weight of 0.9).The other two tasks are used as auxiliary tasks(with a weight of 0.05 respectively)and share the network weight of the three tasks to achieve the goal of improving the main task model.Experiments show that the network sharing of coding layer for multiple tasks with different granularity can im-prove the text feature extraction ability of the model for sentences,so as to improve the emotion recognition accuracy of the model for sentences,especially in fine-grained(micro)emotion classification tasks.