基于多注意力卷积神经网络的特定目标情感分析

发布时间：2018-04-03 02:04

本文选题：注意力机制　切入点：卷积神经网络　出处：《计算机研究与发展》2017年08期

【摘要】：特定目标情感分析作为情感分析一个重要的子任务,近年来得到越来越多研究人员的关注.针对在特定目标情感分析中,将注意力机制和LSTM等序列性输入网络相结合的网络模型训练时间长、且无法对文本进行平行化输入等问题,提出一种基于多注意力卷积神经网络(multi-attention convolution neural networks,MATT-CNN)的特定目标情感分析方法.相比基于注意力机制的LSTM网络,该方法可以接收平行化输入的文本信息,大大降低了网络模型的训练时间.同时,该方法通过结合多种注意力机制有效弥补了仅仅依赖内容层面注意力机制的不足,使模型在不需要例如依存句法分析等外部知识的情况下,获取更深层次的情感特征信息,有效识别不同目标的情感极性.最后在SemEval2014数据集和汽车领域数据集(automotive-domain data,ADD)进行实验,取得了比普通卷积神经网络、基于单注意力机制的卷积神经网络和基于注意力机制的LSTM网络更好的效果.
[Abstract]:As an important sub-task of affective analysis, Goal-specific affective analysis has been paid more and more attention by researchers in recent years.The network model which combines attention mechanism with sequential input network such as LSTM has long training time and can not parallel text input in the analysis of specific target emotion.In this paper, a method for analyzing specific target emotion based on multi-attention convolution neural networks is proposed.Compared with the attention-based LSTM network, the proposed method can receive parallel input text information and greatly reduce the training time of the network model.At the same time, by combining multiple attention mechanisms, the method can effectively compensate for the deficiency of only relying on the content-level attention mechanism, so that the model does not need external knowledge such as dependency parsing.To obtain deeper emotional feature information and identify the emotional polarity of different targets effectively.Finally, experiments are carried out on SemEval2014 data sets and automotive domain data sets. The results are better than those of conventional convolution neural networks, single attention based convolution neural networks and attention based LSTM networks.
【作者单位】：苏州大学计算机科学与技术学院;软件新技术与产业化协同创新中心;符号计算与知识工程教育部重点实验室(吉林大学);
【基金】：国家自然科学基金项目(61272005,61303108,61373094,61472262,61502323,61502329) 江苏省自然科学基金项目(BK2012616) 江苏省高校自然科学研究项目(13KJB520020) 吉林大学符号计算与知识工程教育部重点实验室基金项目(93K172014K04)~~
【分类号】：TP183;TP391.1

【参考文献】