高级检索

    基于双指导注意力网络的属性情感分析模型

    Aspect-Based Sentiment Analysis Model with Bi-Guide Attention Network

    • 摘要: 鉴于深度学习技术的不断发展,越来越多的研究者倾向于使用深度神经网络学习文本特征表示用于情感分析,其中序列模型(sequence models)和图神经网络(graph neural networks)已得到广泛的应用,并取得了不错的效果.然而,对于属性情感分类任务,属性对象与其他单词之间存在远距离的依赖关系,虽然序列型神经网络能捕获句子的上下文语义信息,但是对词语之间的远距离依赖关系无法进行有效学习;而图神经网络虽然可以通过图结构聚合更多的属性依赖信息,但会忽略有序词语间的上下文语义联系.因此结合双向长短时记忆网络(bi-directional long short-term memory, BiLSTM)和图卷积神经网络(graph convolutional network, GCN),提出一种基于双指导注意力网络(bi-guide attention network, BiG-AN)的属性情感分析模型.该模型通过交互指导注意力机制,同时关注到文本的上下文信息和远距离依赖信息,提高了模型对于文本属性级别情感特征的表示学习能力.在4个公开数据集Laptop,Rest14,Rest16,Twitter的实验结果表明,与其他几种基准模型相比,所提模型能够提取到更丰富的属性文本特征,有效提高属性情感分类的结果.

       

      Abstract: Due to the development of deep learning technology, an increasing number of researchers tend to use deep neural network to learn text feature representation for sentiment analysis, where sequence models and graph neural networks have been widely used and achieved good results. However, for aspect based sentiment analysis tasks, there is a long-distance dependency between aspect objects and other words. Although the sequential neural network can capture the contextual semantic information of sentences, the long-distance dependency between words cannot be effectively learned. Graph neural networks can aggregate more aspect-dependent information through graph structures, while ignoring contextual semantic relationships between ordered words. Thus an aspect-based sentiment analysis model named BiG-AN (bi-guide attention network) is proposed. The model combines the advantages of bi-directional long short-term memory (BiLSTM) and graph convolution network (GCN) to capture sentiment features at the aspect level of text, using interactively guiding attention mechanism to focus on contextual and long-distance dependency information in the sentence. The experimental results on four open-source datasets, including Laptop, Rest14, Rest16 and Twitter, show that the proposed model can extract richer aspect-based text features and effectively improve the results of aspect based sentiment classification compared with other benchmark models.

       

    /

    返回文章
    返回