信息工程与人工智能学院知识库

方面级情感分析是一项细粒度的情感分析任务，其目标是判断文本语句中包含的不同方面的情感极性（方面也称为属性）。这种细粒度的情感分析可以更精确的挖掘消费者对产品或服务的评论文本信息，对于管理者来讲，可以将这些文本情感信息作为依据，更准确的改进产品质量或服务态度；而对于其他消费者来讲，也可以借助这些文本情感信息去了解该产品的质量，以此来达到购买产品的目的。

在方面级情感分析任务中，现有的方法大多使用长短时记忆神经网络（Long-Short Memory Neural Network，简称为LSTM）并结合注意力机制的模型，这种模型首先是不需要人为的构建文本特征就可以从文本序列中挖掘到需要的文本信息，其次是结合注意力机制给各文本特征分配不同的权重。但是，LSTM神经网络模型对于较长的文本序列进行训练时，容易出现训练时间较长、产生过拟合现象等问题，而对于引入注意力机制的模型，隐性建模属性与上下文中情感表达的关系，只能获取到文本情感信息而忽略使用语法信息，导致文本特征表示不够全面，使得模型难以正确匹配方面属性和情感属性。因此，针对以上问题，本文提出基于双向门控循环单元(Bidirectional Gate Recurrent Unit，Bi-GRU)融合多种特征信息的神经网络模型（以下简称BRGN模型）。

将Word2vec、Glove、ELMo和BERT四种词向量预训练模型做情感分类对比实验，验证BERT预训练模型构建词向量作为BRGN模型的词嵌入层的合理性。为能够解决LSTM神经网络模型存在的弊端，同时考虑到文本上下文之间的联系，BRGN模型构建双向的GRU神经网络层。词向量通过双向GRU神经网络层，获得词语的语义信息；同时，使用文本序列中词的位置信息，来关注距离方面词位置越近的词对方面情感极性的影响，词向量通过双向GRU神经网络层，获取词语上下文的位置特征，并在位置记忆网络层中存储词的位置信息。利用图卷积神经网络层融入通过句法分析得到的外部知识，能够使模型感知到文本序列的句法结构信息。将文本的句法依存树以邻接矩阵的形式作为图卷积神经网络的输入，接着图卷积神经网络处理邻接矩阵所表示的图结构信息数据，提取相邻接点的特征，将方面词与情感词之间建立关系，更好的进行特征提取。为充分利用记忆网络存储的文本语义信息、位置信息和图卷积神经网络模型得到的句法信息等特征，提出一种考虑到句法等辅助信息的情感注意力层，通过增加当前方面词所对应情感词的权重，筛选出文本序列中与方面词相关的情感词，并辅助情感分类层判断文本情感倾向。

最后，为验证所提出BRGN模型的预测性能，通过在Restaurant、Laptop以及Twitter数据集上的实验结果表明，本文所提出的模型在上述3组数据集上的准确率和F1值都取得了较好的效果，从而证模型的有效性。分析图卷积神经网络层数对模型性能的影响，并通过案列分析，BRGN模型能高效的预测出文本评论中不同方面所属的情感极性，从而验证模型的可靠性。

Aspect-level sentiment analysis is a fine-grained sentiment analysis task whose goal is to determine the sentiment polarity of different aspects (aspects are also called attributes) contained in a text utterance. This kind of fine-grained sentiment analysis can more accurately tap into the textual information of consumers' comments on products or services. For managers, they can use this textual sentiment information as a basis to more accurately improve the quality of products or service attitudes; and for other consumers, they can also use this textual sentiment information to understand the quality of the product, so as to achieve the purpose of purchasing the product.

In aspect-level sentiment analysis tasks, most existing approaches use Long-Short Memory Neural Network (LSTM) combined with an attention mechanism model, which firstly, does not require artificial construction of text features to mine the required text information from text sequences, and secondly, combines This model firstly, does not need to construct text features artificially to mine the required text information from text sequences, and secondly, combines the attention mechanism to assign different weights to text features. However, the LSTM neural network model is prone to problems such as long training time and overfitting when training longer text sequences, and for the model that introduces the attention mechanism, it implicitly models the relationship between attributes and emotional expressions in the context, and can only obtain textual emotional information while ignoring the use of grammatical information, which leads to a less comprehensive representation of text features and makes it difficult for the model to correctly match aspects attributes and sentiment attributes. Therefore, to address the above problems, this paper proposes a neural network model based on Bi-GRU fusing multiple types of information (BRGN Model).

The word vectors obtained from four pre-training models, namely Word2vec, Glove, ELMo and BERT, are used for sentiment classification comparison experiments to verify the rationality of the BERT pre-training model to construct suitable word vectors as the word embedding layer of the BRGN model. In order to solve the drawbacks of the LSTM neural network model, while taking into account the connection between text contexts, the BRGN model builds bidirectional GRU neural network layers. The word vector obtains the semantic information of words through the bidirectional GRU neural network layer; meanwhile, the position information of words in the text sequence is used to focus on the influence of words closer to the position of aspectual words on the emotional polarity of aspectual words. The word vector obtains the position characteristics of word context through the bidirectional GRU neural network layer and stores the position information of words in the position memory layer.The incorporation of external knowledge obtained through syntactic analysis using the graph convolutional neural network layer enables the model to perceive the syntactic structural information of text sequences. The syntactic dependency tree of the text in the form of an adjacency matrix is used as the input to the graph convolutional neural network, and then the graph convolutional neural network processes the graph structure information data represented by the adjacency matrix, extracts the features of the adjacent contacts, and establishes relationships between aspect words and sentiment words, thus assisting the model in aspect-level sentiment analysis. In order to make full use of the features such as semantic information of text stored in the memory network, location information and syntactic information obtained from the graph convolutional neural network model, this paper proposes a sentiment attention layer that takes into account auxiliary information such as syntax. By increasing the weight of the lexical vector representation of the sentiment word corresponding to the current aspect word, the sentiment words related to the aspect word in the text sequence are filtered, and the sentiment classification layer is assisted to determine the sentiment classification layer is used to determine the sentiment tendency of the text.

Finally, in order to verify the prediction performance of the proposed BRGN model, the experimental results on Restaurant, Laptop and Twitter datasets show that the proposed model achieves better accuracy and F1 values on the above three datasets, thus confirming the effectiveness of the model. The effects of the number of layers of the graph convolutional neural network on the performance of the model are also verified, and the reliability of the model is confirmed by the example analysis that the BRGN model proposed in this paper can efficiently predict the sentiment polarity belonging to different aspects in text comments.