子句级别的自注意力机制的情感原因抽取模型

Emotion cause extraction model based on clause self-attention mechanism

情感原因抽取(ECE)是情感分析领域的一项重要子任务,旨在识别给定文档中某种情绪表达所对应的原因.现有的一些工作将该任务定义为子句分类任务,关注了文档和子句之间的联系,而忽略情感描述子句与情感原因子句的直接语义联系,同时存在标签不平衡问题,使得情感原因子句位置难以定位.因此,提出了一个基于子句的自注意力机制同时结合了子句相对位置关系的神经网络模型去寻找情感原因子句.为了更加突显句子的局部特征,利用卷积神经网络抽取每个子句的上下文特征.模型首先使用双向长短期记忆网络编码子句信息,融合子句位置特征后,利用自注意力机制计算情感原因子句和情感描述子句之间的语义信息,并结合子句局部上下文特征,抽取情感原因子句.在基于新浪城市新闻的情感原因抽取中文数据集上,查全率R达到83.83%,优于目前的基线方法.

Emotion cause extraction(ECE)is an important subtask in the field of emotion analysis,which aims to identify the reasons for a certain emotional expression in a given document.Some existing works define this task as a clause classification task,focusing on the connection between the document and the clause,while ignore the direct semantic connection between the emotional description clause and the emotional cause clause.At the same time,label imbalance makes the clause position more difficult to locate.Therefore,this paper proposes a neural network model with a self-attention mechanism based on clauses that combines the relative positions of clauses to find emotional cause clauses.In order to more highlight the local features of the clause,the convolutional neural network is used to extract the context information of each clause.The model first uses the bidirectional long-term and short-term memory network to encode the clause information.After fusing the position features of the clauses,the self-attention mechanism is used to calculate the semantic information between the emotional cause clause and the emotional description clause.Finally,the contextual feature of the clause is combined with above information by the model to extract the emotional reason clause.On experiment of Chinese datasets based on Sina City News,recall reached 83.83%,which outperforms the state-of-the-art baseline methods.