基于BERT的多信息融合方面级情感分析模型

Aspect-Level Sentiment Analysis Model based on BERT with Multi-Information Fusion

方面级情感分析任务旨在判断文本语句不同方面的情感极性,是自然语言处理领域的热点任务之一。当前基于BERT的方面级情感分析方法大多仅将其作为预训练词嵌入工具,没有充分利用其本身的语义提取能力和任务处理能力,在利用句法及方面词位置等信息时,常忽略了不同外部信息的关联性。针对上述问题,提出一种基于BERT的多信息融合网络模型。首先,构建基于BERT词嵌入的主路径和辅路径,辅路径可以避免对词嵌入特征产生干扰;其次,在辅路径下,根据文本语句的句法依存树和多信息融合算法让不同词语与方面词的句法距离、位置距离等信息产生交互;最后,使用卷积神经网络和注意力机制将主路径与辅路径的特征融合。3个数据集的实验结果表明,提出的模型是有效的。

Aspect-level sentiment analysis aims to determine the sentiment polarity of different aspects of a text sentence,which is one of the hot tasks in the field of natural language processing.Most of the current BERT-based aspect-level sentiment analysis methods only use it as a pre-trained word embedding tool,without making full use of its own semantic extraction capabilities and task processing capabilities,and often ignore the correlation of different external information when using information such as syntax and aspect word position.To solve the above problems,this paper proposes a multi-information fusion network model based on BERT.Firstly,the main path and auxiliary path based on BERT word embedding are constructed,and the auxiliary path could avoid interference with the word embedding features.Secondly,under the auxiliary path,the syntactic distance,position distance and other information between different words and as-pect words are interacted according to the syntactic dependency tree of the text sentence and the multi-information fusion algorithm.Finally,convolutional neural network and attention mechanism are used to fuse the features of the main path and the auxiliary path.Experimental results on three datasets show that the proposed model is effective.