Aspect Sentiment Triplet Extraction (ASTE) aims to jointly identify aspect terms, opinion expressions, and sentiment polarities from user reviews. Most existing ASTE models are developed on benchmark datasets containing short sentences and explicit aspect–opinion pairs, and thus generalize poorly to realistic reviews that contain implicit aspects and long, compositional opinion spans. This paper investigates ASTE under these challenging conditions on the diversified multi-domain DMASTE benchmark. We propose IBR, a generative framework with implicit reasoning and boundary-aware regularization for ASTE. IBR integrates two complementary modules: the Implicit Triplet Constructor with Explicit Structure (ITC-ES), which treats explicit sentences as structural templates and performs minimal edits on both inputs and outputs to construct pseudo-implicit training samples with explicit structure, while an explicit–implicit classifier guides the model to infer missing aspect terms; and the Syntax-Guided Opinion Boundary (SG-OB) module, which introduces syntax-guided opinion-boundary supervision during training to improve the localization of long and syntactically complex opinion expressions. Experiments on DMASTE show that IBR achieves consistent gains over reported baselines across in-domain and transfer settings. Further analyses suggest that ITC-ES mainly benefits implicit triplet extraction, while SG-OB improves opinion boundary localization in long and compositional expressions.
Jiang et al. (Mon,) studied this question.