Template-Order Driven Feature Integration With Generative Models for Aspect Sentiment Triplet Extraction

Aspect-based sentiment analysis (ABSA), which explores the nuanced sentiments individuals express toward specific services or products, has shown significant potential in practical applications. Recently, the aspect sentiment triplet extraction field has witnessed significant advancements through the prowess of generative models. However, existing methods face critical limitations: (1) fixed-order decoding (e.g., aspect term <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$\rightarrow$</tex-math></inline-formula> opinion term <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$\rightarrow$</tex-math></inline-formula> sentiment polarity) ignores interdependencies between elements; (2) token-by-token generation fails to model term boundaries, which struggles with multi-word terms or sentences containing multiple triplets. To address these challenges, we propose a Template-Order driven Feature Integration (TOFI) framework, which integrates two novel modules: Template-Order Prompt (TOP) and Feature Progressive Sequence Labeling (FPSL). TOP employs various orders of sentiment elements to prompt the model to generate sentiment tuples from different perspectives, each using a different element order. It then aggregates the sentiment tuples that appear across all orders. FPSL leverages sequence labeling to integrate the rich semantic information of aspect terms and opinion terms into the model, enhancing its ability to capture semantic nuances. Furthermore, the TOFI framework introduces unified multi-domain training, prefixing domain identifiers to inputs for multi-task learning across datasets. Extensive experiments on multiple public benchmarks demonstrate that the proposed framework consistently surpassed representative baseline models.