信息工程与人工智能学院知识库

随着互联网应用的快速发展，各大社交平台的兴起，人们越来越喜欢在互联网上发表自己的观点、意见或表达情绪等。通过收集、整理并分析这些带有用户情感倾向的评论文本，商家可以更好地了解客户的行为和偏好，政府也可以及时了解群众对某项政策的看法，并适当引导公众舆论，具有一定的社会和商业价值。近些年流行的方面级情感分析是先提取方面实体，再去判断该方面实体所对应的情感态度，以一种更加细致的方式来判断某一方面实体所蕴含的情感色彩，并且当句中有多个方面时，能够将不同方面与其情感极性一一匹配，因其能够满足更多的实际需要而受到广泛关注。

目前，大多数基于方面级的情感分析模型是依靠依存句法树获取句法信息，并结合神经网络来进行情感极性的预测，但是仅根据依赖树获取信息一方面无法区分不同上下文词的重要程度，导致语义信息提取不充分；另一方面会因为单独计算注意力而忽略上下文信息与目标方面之间的联系，使一些重要特征信息丢失。因此，为解决上述问题，本文提出了基于依存句法树的交互式方面级情感分析模型（An Interactive Aspect-Level Sentiment Analysis Model Based on a Dependency Syntax Tree，以下简称为SIGAT模型）。本模型主要包括两个部分：一部分是基于句法注意机制的目标方面提取，该模块主要是将语法信息与注意力机制相结合。首先利用语法解析器生成依存句法树，随后使用句法注意机制对依存树上的语法信息进行编码，建模上下文词的语义关系，并有选择性的关注语法路径上接近目标方面的语境词，对句法相对距离近的意见词赋予更多的关注，同时引入高斯函数降低计算的复杂度，避免因权重急剧下降导致情感信息的丢失。另一部分是基于图注意力网络的上下文特征提取，以依存句法树为辅助，不仅考虑句法信息，也考虑单词之间的依赖关系，丰富单词的表示；最后，为加强二者之间的联系，通过交互注意力机制使基于图注意力网络的上下文特征与基于句法注意机制的目标方面特征交互学习，用对方的信息来补充自身的特征信息，提高特征表达能力，经特征融合后，输入至全连接层分类输出结果，以判断目标方面的情感态度。

最后，为证明本文所提出的SIGAT模型的分类效果，在三个领域的数据集:SemEval-2014 Task 4中的餐厅评论（Restaurant）、笔记本电脑评论（Laptop）以及ACL-14 Twitter上进行实验。通过与其他基线模型进行对比，SIGAT模型在数据集上的准确率（Acc）和F1值优于其他模型，证实了该模型的可靠性。同时，通过消融实验和对不同影响因素的分析发现，SIGAT模型中的依存句法信息和交互注意力机制在提高模型性能方面有重要贡献。

With the rapid development of Internet applications and the rise of major social networking platforms, people are increasingly interested in expressing their views, opinions, etc. on the Internet. By collecting, collating, and analyzing these comment texts with users' sentiment tendencies, businesses can better understand their customers' behavior and preferences, and governments can keep abreast of the public's views on a certain policy and appropriately guide public opinion, which has certain social and commercial value. Aspect-Based Sentiment Analysis , which has become popular in recent years, extracts the aspect entities first and then determines the sentiment attitudes corresponding to them, in a more detailed way, and when a sentence contains multiple aspects, it is possible to match the sentiment polarities of different aspect entities according to their The ability to match different aspectual entities to their corresponding emotional polarity when multiple aspects are included in a sentence is of great interest because of its ability to meet a wider range of practical needs.

Currently, most aspect-level based sentiment analysis models rely on dependency syntactic trees to obtain syntactic information and combine with neural networks for sentiment polarity prediction, but obtaining information based on dependency trees alone cannot distinguish the importance of different contextual words on the one hand, resulting in inadequate extraction of semantic information; on the other hand, the connection between contextual information and the target aspect will be ignored because attention is calculated separately, making some models inefficient. On the other hand, some important feature information is lost because the contextual information and the target aspect are neglected by calculating attention alone. Therefore, to solve the above problems, this paper proposes an Interactive Aspect-Level Sentiment Analysis Model Based on a Dependency Syntax Tree (hereafter referred to as the SIGAT model). The model consists of two main parts: one is the extraction of target aspects based on syntactic attention mechanisms. The module mainly combines syntactic information with attention mechanisms, first using a syntactic parser to generate a dependency syntax tree, then using syntactic attention mechanisms to encode the syntactic information on the dependency tree, modeling the semantic relations of contextual words, and selectively paying attention to contextual words close to the target aspect on the syntactic path, giving more attention to opinion words close to the syntactic relative distance, and introducing a Gaussian function to reduce the complexity of the computation and avoid the loss of sentiment information due to a sharp drop in weighting. Another part is the graph attention network based contextual feature extraction, aided by a dependency syntactic tree, which not only considers syntactic information but also the dependency relationship between words, enriching the representation of words; and finally, through the interactive attention mechanism, the contextual features based on the graph attention network and the target aspect features based on the syntactic attention mechanism are learned interactively, using each other's information to supplement Finally, through the interactive attention mechanism, the contextual features based on the graph attention network and the target aspect features based on the syntactic attention mechanism are learned interactively, using each other's information to supplement their own feature information and improve the expression of the features.

Finally, to demonstrate the classification effectiveness of the SIGAT model proposed in this paper, experiments were conducted on three domain datasets:Restaurant reviews (Restaurant), Laptop reviews (Laptop) and ACL-14 Twitter in SemEval-2014 Task 4. By comparing with other baseline models, the SIGAT model outperformed the accuracy (Acc) and F1 values on the dataset, confirming the reliability of the model. Also, the ablation experiments and the analysis of different influencing factors revealed that the syntactic information and the interaction attention mechanism on the dependency tree in the SIGAT model made an important contribution to improving the model's performance.