基于多任务融合模型的情感原因配对

Multi-Task Fusion Model-Based Emotion Cause Pair Extraction

情感原因配对任务的目标是自动抽取文本的情感、原因和配对关系,与情感原因发现相比更具有挑战性.针对现有的情感原因配对方法容易忽视情感和原因之间的双向依赖问题,提出将情感原因配对任务分解为子句级别的序列标注任务和标签配对任务.首先,借助预训练语言模型、双向长短时记忆网络和注意机制获取子句向量;然后,采用卷积神经网络捕捉句子间的相邻信息,结合条件随机场进行序列标注;最后,采用基于规则的方式完成情感原因配对.实验结果表明,在情感原因匹配任务上,精度为0.726、召回率为0.688、F1为0.703,验证了该方法的有效性.

The goal of Emotion Cause Pair Extraction task is to automatically extract emotional reasons and their corresponding relationships from text.Therefore,which is more challenging than Emotion Cause Extraction.The existing methods for Emotion Cause Pair Extraction in text often overlook the bidirectional dependency between emotions and causes.To address this issue,we propose decomposing Emotion Cause Pair Extraction task into sub-sentence level sequence labeling and label pairing tasks.Firstly,we use pre-trained language models,bidirectional long short-term memory networks,and attention mechanisms to obtain sub-sentence vectors.Then,we adopt convolutional neural networks to capture adjacent information between sentences,and combine it with a conditional random field for sequence labeling.Finally,we use rule-based methods to complete the emotion-cause pairing.The experimental results show that the precision for emotional clause prediction is 0.726,the recall is 0.688,and the F1 score is 0.703 for the emotion-cause pair matching task,thus validating the effectiveness of the proposed method.