Aspect-Level Sentiment Analysis Based on the Convolutional Neural Network and Positional Enhancement Vectors
Qianquan Wang, Lin Zhang · 2024
The Aspect-Lased Sentiment Analysis (ABSA) is able to effectively extract key information helpful for decision-making from a large amount of user-generated content currently available on various Internet platforms by predicting the sentiment polarity of a given aspect word in a sentence. Existing studies usually ignore the positional information of the context on the relative distance of aspectual words, resulting in models that may focus on irrelevant words. In this paper, we hypothesize that the positional weight of the context (i.e., words closer to the aspect word are more important) has a positive effect on the sentiment bias judgments of aspectual words, and we propose a mechanism for capturing the positional information, i.e., positional weight encoding, to further improve the prediction accuracy of the affective tendencies corresponding to aspectual words. In this paper, the dynamic word vector representation is obtained using the BERT model and fused by the proposed positional weights to obtain the positionally enhanced word vector representation, and then the textual semantic features are extracted using dual-channel networks. Experiments show that the accuracy and F1 value of the proposed model in this paper is significantly improved on the SemEval 2014 Task 4 dataset, confirming the lifting effect of the model we studied.