Integrating Multi‐Scale Gated Convolution and Aspect‐Aware Bidirectional Mamba for Aspect Sentiment Triplet Extraction

ABSTRACT Serving as a subtask of aspect‐based sentiment analysis, aspect sentiment triplet extraction (ASTE) enables the extraction of triplets that consist of aspect terms, opinion terms, and sentiment polarities from user reviews. This enables the mining of user preferences and supports merchants in optimizing products and services. However, most existing studies suffer from limitations such as insufficient ability to capture fine‐grained sentiments and failure to effectively mine long‐range dependencies. To address these limitations, this paper proposes an ASTE model that integrates multi‐scale gated convolution and aspect‐aware bidirectional Mamba. Specifically, the model designs a multi‐scale gated convolution module, which uses a gating mechanism to dynamically adjust weights and leverages multi‐scale convolutional layers to enhance the model's capability of capturing fine‐grained sentiment features. Subsequently, aspect‐aware bidirectional Mamba is introduced, which explicitly focuses on aspect‐related information through aspect‐aware gating and adopts a bidirectional structure to capture contextual dependencies, thereby effectively capturing long‐range dependencies. Then, an adaptive feature fusion module is constructed to aggregate valid information. Results from experiments conducted on the four sub‐datasets of ASTE‐DATA‐V2 reveal that the proposed model outperforms baseline models, thus comprehensively verifying its validity.