信息工程与人工智能学院知识库

自然语言资源是一种以文本形式呈现的资源，其背后往往隐藏着人们对某种事件或商品的看法。如果我们能够对这些资源进行准确分析，找出隐藏在文字背后的情感态度，并对这些数据进行合理高效的利用，将会给企业、消费者以及相关部门带来巨大的利益。文档或句子层面的情感分析任务是对整个文档或句子进行情感分析。然而，由于一篇文章或句子中存在多个具有不同情感极性的主体，这种情感分析方法的准确性就会降低。因此，更细粒度方面的情感分析引起了越来越多的关注。基于此，人们提出了许多方法，例如用于人工设计特征的支持向量机、基于注意力机制的长短期记忆网络、基于依赖解析的图神经网络。虽然这些方法都有很不错的性能，但他们都忽略了一项重要的语法信息：依赖标签。

回顾现有的 ALSA 方法，从最初基于手工标注特征工程的方法，到后来利用深度学习，人们开始使用 LSTM 的各种变体来解决这个问题。随着注意力和最近GNN 的引入来解析依赖关系的语法特征，方面级情感分析的模型性能正在逐步增强。模型在简单的语义特征提取上已经很难改进了。但在语法特征的提取上，还有很大的改进空间。通过观察当前在语法特征提取方面性能良好的模型，发现这些模型只是简单地提取依赖弧的连接信息作为语法特征并将其输入到 GCN 或GAT 当中。但除了依赖弧的连接信息外，“直接宾语”“形容词修饰语”等依赖弧的标签也是提供句法信息非常重要的部分。所以模型可以将这些依赖弧标签转化为嵌入，对句子进行注意力操作，以捕获这句话中的重要部分。此外，考虑到依赖树中单词和依赖弧标签之间的关系，本文使用图神经网络来处理单词嵌入和依赖标签嵌入。 使用预训练的BERT 模型作为编码器，带来了巨大的性能提升。

基于上述想法，本文提出了两个有效的模型来完成方面级别的情感分类任务。一种是基于 Bi-LSTM,一种是基于 BERT。为使模型的性能达到最优，选取了具有代表性的模型在三个公开数据集上进行了比较实验。实验结果显示，本文提出的模型在预测准确率和 F1 值上有不同幅度的提升。为了进一步了解模型中注意力层对模型性能的影响，设置了一个删除了依赖嵌入层、依赖 GCN 和依赖注意力模块的 Non-Dep 模型以及改变了两个注意力模块的运行方式来验证注意力层对模型性能的提升作用。最后进行实例研究模型的推理过程和两个注意力模块的作用。

Natural language resources are a type of resource presented in textual form, often hiding people's views on a certain event or commodity. If we can accurately analyze these resources, identify the emotional attitudes hidden behind the text, and make reasonable and efficient use of this data, it will bring huge benefits to enterprises, consumers, and relevant departments. The task of sentiment analysis at the document or sentence level is to perform sentiment analysis on the entire document or sentence. However, due to the presence of multiple subjects with different emotional polarities in an article or sentence, the accuracy of this sentiment analysis method will be reduced. Therefore, more fine-grained emotional analysis has attracted increasing attention. Based on this, many methods have been proposed, such as support vector machines for manually designing features, long short-term memory networks based on attention mechanisms, and graph neural networks based on dependency analysis. Although these methods all have good performance, they all overlook an important syntactic information: dependency labels.

Looking back at the existing ALSA methods, from the initial approach based on manually annotated feature engineering to the later use of deep learning, people began to use various variants of LSTM to solve this problem. With the introduction of attention and recent GNN to parse the syntactic features of dependency relationships, the performance of aspect level sentiment analysis models is gradually improving. The model is already difficult to improve on simple semantic feature extraction.

But there is still a lot of room for improvement in the extraction of grammatical features. By observing the current models that perform well in syntax feature extraction, it is found that these models simply extract the connection information of dependency arcs as syntax features and input them into GCN or GAT. But in addition to the connection information of the dependent arc, labels such as "direct object" and "adjective modifier" are also very important parts in providing syntactic information. So the model can convert these dependency arc labels into embeddings, perform attention operations on the sentence, and capture the important parts of the sentence. In addition, considering the relationship between words and dependency arc labels in the dependency tree, I use graph neural networks to handle word embedding and dependency label embedding. The use of pre trained BERT models as encoders has brought significant performance improvements.

Based on the above ideas, this article proposes two effective models to complete aspect level sentiment classification tasks. One is based on Bi-LSTM, and the other is based on BERT. To achieve optimal performance of the model, representative models were selected for comparative experiments on three public datasets. The experimental results show that the proposed model has different degrees of improvement in prediction aaccuracy and F1 value. In order to further understand the performance of the attention layer in the model, a Non Dep model was set up, which removed the dependency embedding layer, dependency GCN, and dependency attention module, and changed the operation mode of the two attention modules to verify the improvement effect of the attention layer on model performance. Conduct case study on the reasoning process of the model and the role of two attention modules.