一种循环卷积注意力模型的文本情感分类方法  

Recurrent Convolution Attention Model for Sentiment Classification

作　　者：陈千[1,2] 车苗苗 郭鑫 王素格[1,2] CHEN Qian;CHE Miao-miao;GUO Xin;WANG Su-ge(School of Computer&Information Technology,Shanxi University,Taiyuan 030006,China;Key Laboratory of Computational Intelligence and Chinese Information Processing,Ministry of Education,Taiyuan 030006,China)

机构地区：[1]山西大学计算机与信息技术学院,太原030006 [2]山西大学计算智能与中文信息处理教育部重点实验室,太原030006

出　　处：《计算机科学》2021年第2期245-249,共5页Computer Science

基　　金：国家自然科学基金(61502288,61403238);山西省基础研究计划项目(201901D111032,201701D221101);山西省重点研发计划项目(201803D421024)。

摘　　要：情感分类对推荐系统、自动问答、阅读理解等下游应用具有重要应用价值,是自然语言处理领域的重要研究方向。情感分类任务直接依赖于上下文,包括全局和局部信息,而现有的神经网络模型无法同时捕获上下文局部信息和全局信息。文中针对单标记和多标记情感分类任务,提出一种循环卷积注意力模型(LSTM-CNN-ATT,LCA)。该模型利用注意力机制融合卷积神经网络(Convolutional Neural Network,CNN)的局部信息提取能力和循环神经网络(Recurrent Neural Network,RNN)的全局信息提取能力,包括词嵌入层、上下文表示层、卷积层和注意力层。对于多标记情感分类任务,在注意力层上附加主题信息,进一步指导多标记情感倾向的精确提取。在两个单标记数据集上的F1指标达到82.1%,与前沿单标记模型相当;在两个多标记数据集上,小数据集实验结果接近基准模型,大数据集上的F1指标达到78.38%,超过前沿模型,表明LCA模型具有较高的稳定性和较强的通用性。Sentiment classification has important application value for downstream applications,including recommendation system,automatic question answering and reading comprehension.It is an important research direction in the field of natural language processing.The task of sentiment classification depends on global and local information hidden in context.However,exis-ting neural network models can not capture the local and global information of context at the same time.In this paper,a recurrent convolutional attention model(LSTM-CNN-ATT,LCA)is proposed for single label and multi-label sentiment classification tasks.It uses attention mechanism to fuse the local information extraction ability of convolutional neural network and the global information extraction ability of recurrent neural network,including word embedding layer,context representation layer,convolution layer and attention layer.For the multi-label sentiment classification task,the topic information is added to the attention layer to further guide the accurate extraction of multi-label emotion tendency.The F1 index on two single label datasets reaches 82.1%,which is equivalent to the frontier single label model.On two multi-label datasets,the experimental results on small datasets are close to the benchmark model,and the F1 index on large datasets reaches 78.38%,which is higher than the state-of-the-art model.It indicates that LCA model has high stability and strong universality.

关 键 词：循环神经网络 卷积神经网络 注意力机制 情感分类 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]