信息工程与人工智能学院知识库

方面级情感分类是情感分析中的细粒度任务，其目标是识别句子中某一方面对应的情感态度，这类对针对具体方面的情感提取能够在个人、企业和相关部门作出决策时提供更加精准的数据支撑，因此具有非常大的现实意义。目前大多数情感分类模型在构建时往往只提取文本信息的某一类特征，而忽略了将多种特征融合也能给模型带来效果的提升。为解决以上问题，本文提出一种基于双通道信息融合的GCNN模型。模型构建主要思路如下：

（1）在方面级情感分类任务中，传统情感分类模型往往将词语作为单一嵌入，而忽略词语的词性信息和位置信息，因此在模型预测表现上并不理想。本文用BERT预训练模型增强传统词嵌入的情感语义表征能力，同时将词性信息和位置信息融入得到最终的词向量表示，缓解词向量表征不足的问题。

（2）卷积操作能够实现文本特征的有效提取，本文设计了基于多粒度卷积操作的MulCNN模块用于提取文本的语义特征。不同于常用的单个卷积核，模块能同时采用三个不同粒度的卷积核提取文本序列的N-gram特征，粒度不同的卷积核提取到的特征矩阵并不相同，相较于单粒度卷积操作，本文提出的MulCNN模块能够提取到句子中更丰富的表征信息。

（3）尽管N-gram特征能够很好的帮助模型判断方面词的情感极性，但是单一特征的深度学习模型并未考虑到句法结构的重要性，本文设计了基于句法关系的DepGCN模块。模块首先通过依赖解析模型学习句子依赖关系得到词语间的逻辑距离，接着以词语和依赖关系分别作为图卷积神经网络的初始节点和边，得到包含丰富句法特征的向量表示。

（4）传统模型为融合特征往往对特征向量进行简单的拼接操作，但是这类操作并没有进一步挖掘信息间的关联性。本文同时兼顾语法特征相关性和语义特征互补性，基于注意力机制构建了特征融合层用于双通道的特征融合，达到了增强特征向量表达的能力。

经过BERT预训练后的词嵌入表示蕴含着更为丰富的特征信息，这些高质量的词向量经过双通道架构能够同时学习句子的语义和语法特征，并在特征融合层实现特征的有效交互，充分发挥语义和语法特征的互补优势，增强了模型的鲁棒性。当句子意思表达充分时，GCNN模型可以依靠包含语义信息的MulCNN模块实现情感分类；当句子句法关系明显时，GCNN模型可以依靠包含句法信息的DepCNN模块实现情感分类。通过将GCNN模型与选取的12个代表模型在三个公开数据集上做对比实验可知，本文提出的模型在预测准确率和F1值方面表现出色，同时本文还设置了一系列实验证明本文提出的GCNN模型能够有效挖掘并融合特征信息，证实了GCNN模型的可靠性和有效性。

Aspect-level sentiment classification is a fine-grained task in sentiment analysis, and its goal is to identify the sentiment attitude corresponding to a certain aspect of a sentence. This kind of aspect-specific sentiment extraction can provide more accurate data to support individuals, enterprises, and related departments in making decisions, and therefore has great practical significance. Most of the current sentiment classification models are constructed by extracting only one type of feature of text information, ignoring the fact that the fusion of multiple features can also improve the model's effectiveness. To solve the above problems, this paper proposes a GCNN model based on two-channel information fusion. The main idea of model construction is as follows:

 (1) In aspect-level sentiment analysis tasks, traditional sentiment classification models tend to treat words as a single embedding and ignore lexically and location information of words, and thus are not satisfactory in model prediction performance. In this paper, we use a BERT pre-training model to enhance the sentiment-semantic representation of traditional word embeddings, while incorporating lexical and location information to obtain the final word vector representation to alleviate the problem of insufficient word vector representation.

 (2) In this paper, we design a MulCNN module based on multi-granularity convolutional operations to extract semantic features of the text. Unlike the commonly used single convolutional kernel, the module can use three different granularity convolutional kernels to extract N-gram features of text sequences at the same time, and the feature matrices extracted by the different granularity convolutional kernels are not the same.

(3) Although N-gram features can well help the model to judge the sentiment polarity of aspect words, the deep learning model with single features does not consider the importance of syntactic structure, and the DepGCN module based on syntactic dependencies is designed in this paper. The module first learns the sentence dependency relationship to get the logical distance between words by the dependency resolution model, and then uses words and dependencies as the initial nodes and edges of the graph convolutional neural network, respectively, to get a vector representation containing rich syntactic features.

(4) Feature information is the key to the model's ability to accurately predict sentiment polarity. Traditional models often perform simple splicing operations on feature vectors to fuse features, but such operations do not further explore the correlation between information. In this paper, we take into account both syntactic feature relevance and semantic feature complementarity and build an information fusion layer for two-channel feature fusion based on the attention mechanism, which achieves the ability to enhance feature vector representation.

The word embedding representation after BERT pre-training contains richer feature information, and these high-quality word vectors can learn both semantic and syntactic features of the sentences through the dual-channel architecture, and realize effective feature interaction in the feature fusion layer to give full play to the complementary advantages of semantic and syntactic features and enhance the robustness of the model. When the sentence meaning is adequately expressed, the GCNN model can rely on the MulCNN module containing semantic information to achieve sentiment classification; when the syntactic relationship of the sentence is obvious, the GCNN model can rely on the DepCNN module containing syntactic information to achieve sentiment classification. By comparing the GCNN model with the selected 12 representative models on three public datasets, it can be seen that the model proposed in this paper performs well in terms of prediction accuracy and F1 value, and a series of experiments are also set up in this paper to prove that the GCNN model proposed in this paper can effectively mine and fuse feature information, which confirms the reliability and validity of the GCNN model