Enhancing language feature embeddings with learnable scaling for aspect sentiment triplet extraction

Aspect-Based Sentiment Analysis (ABSA) seeks to detect the emotional information conveyed by users regarding specific features in text, encompassing various components such as aspect terms, opinion words, and sentiment orientation. Existing studies usually use an end-to-end approach to fuse multiple linguistic features without dynamically adjusting the contribution of each feature, which limits the model's performance capability in complex tasks. To tackle this problem, this paper introduces a Learnable Scaling-based Graph Convolutional Network(LS-GCN). This model enables the learnable adjustment of linguistic features, thereby boosting its ability to adapt to diverse linguistic characteristics, which in turn enhances the effectiveness of sentiment triad extraction.The model encodes linguistic information such as semantic, syntactic, and positional information of the text into multidimensional feature vectors by capturing the probability distribution of the relationship between word pairs through the bimodal attention module, and then introduces a learnable scaling module for linguistic features to dynamically adjust the weights of different linguistic features. Finally, a convolutional operation is used to achieve multi-feature fusion, which ultimately achieves the joint extraction of affective triplets. The experimental results of the evaluation based on public datasets show that the proposed LS-GCN model effectively mitigates the problems of insufficient feature scaling and insufficient language feature collaboration, and improves the information interaction and fusion effect between different language features. It also outperforms existing benchmark models in the triplet extraction task.