Adaptation Blocks-Based Semi-Siamese Inference Network for Aspect Sentiment Triplet Extraction

Aspect Sentiment Triplet Extraction (ASTE) stands as one of the most challenging tasks in fine-grained sentiment analysis. However, most models struggle to determine the sentiment categories of multiple words within complex contexts. We propose a novel approach called Adaptive Block-based SemiSiamese Inference Network (ABSIN) to address these challenges. This method decomposes sentences into multiple adaptive blocks for sentiment triplet extraction, where each block may contain multiple words corresponding to different roles. We employ a Semi-Siamese Inference Network combining greedy and selective reasoning to extract the longest aspect/view with the highest label probability. The computational process is constrained by balancing the loss function. Our model outperforms existing state-of-the-art approaches across public datasets.