结合多种注意力机制的方面词提取方法

方面词提取是方面级情感分析中最重要的子任务之一,其旨在从评论文本中找出意见目标。当前对于方面词提取主要使用卷积神经网络(Convolutional Neural Networks, CNN)和双嵌入的方法,但传统的CNN模型受限于卷积核感受野,不能很好地获取全局信息。为此,该文提出了一种基于双嵌入和多种注意力的方面词提取模型。联合使用non-local网络能够更好地捕获长范围依赖关系,使用与跳跃连接相结合的空间注意力能够更好地捕获文本的字符特征。该文模型在Laptop数据集和Restaurant数据集上分别进行了实验,F₁值分别为83.39%和76.26%。与多个基线模型相比,该文提出的模型性能更优。

Aspect Terms Extraction Based on Multiple Attention Mechanisms

Aspect terms extraction is the task of extracting the entity attributes commented in the opinions, which is a sub-task of aspect-based sentiment analysis. To improve the current solution base on convolutional neural networks (CNN), this paper proposes an aspect terms extraction model based on double embedding and multiple attentions. It better captures the long-range dependencies by a joint Non-local networks, betters capture the characters in the text by the spatial attention combined with jump connections. Experimented on the laptop dataset and the restaurant dataset, the proposed method achieves F1 values of 83.39% and 76.26%, respectively, better than multiple baseline models.