Title: Optimising of dynamic aspect attention for sequence modelling via reinforcement learning

Authors: P.M. Diaz; M. Julie Emerald Jiju

Addresses: Department of Mechanical Engineering, Ponjesly College of Engineering, Nagercoil, Tamil Nadu 629003, India ' Department of MCA, ‎CSI Institute of Technology, Thovalai, Kanyakumari, Tamil Nadu, India

Abstract: Aspect-based sentiment classification represents a sophisticated approach within sentiment analysis, focusing on identifying sentiment polarity towards specific aspects or entities mentioned in the text. Despite the advancements in ABSC techniques, dynamically selecting and re-weighting crucial words in a sentence remains a challenging task. To solve this issue, this research proposes a novel methodology named DRA-SAC, which combines the dynamic reweighting adapter with soft actor-critic, to address these challenges. The DRA component dynamically selects and re-weights important words within a sentence, leveraging an attention mechanism to focus attention on critical aspects of the context. By adapting attention weights, DRA enhances the model's ability to capture aspect-specific sentiment information. Subsequently, SAC guides the learning process, facilitating the adjustment of attention weights through reinforcement learning techniques. The performance of the proposed model is evaluated on four benchmark datasets Twitter US airline, Flipkart product review, laptop, and restaurant dataset. The experimental results demonstrate that the proposed method achieves 99% accuracy compared to the existing methods.

Keywords: aspect based sentiment classification; dynamic reweighting adapter; DRA; soft actor-critic; SAC; reinforcement learning; BERT.

DOI: 10.1504/IJIEI.2026.154016

International Journal of Intelligent Engineering Informatics, 2026 Vol.14 No.2, pp.129 - 149

Received: 21 Jun 2024
Accepted: 20 Sep 2024
Published online: 10 Jun 2026 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article