Quantitative Reflection for Mental State Analysis via Aspect Sentiment Triplet Extraction

Aspect Sentiment Triplet Extraction (ASTE) is an important technique for mental state analysis, identifying aspects, opinions, and their sentiment polarities from user-generated content. A fundamental challenge lies in capturing the semantic dependencies across different sentiment elements. Most existing methods are model-centric, primarily modeling tripletlevel semantics by aligning predictions with the ground truth. However, users freely post content with varying mental states, showing informal linguistic styles. Such methods may struggle to generalize to challenging cases with element-level diversity. This paper introduces Metamorphic Relation Guided Quantitative Reflection (MaRS), a novel data-centric paradigm that leverages generated testing counterparts of existing samples to improve ASTE. Particularly, inspired by software metamorphic testing, we construct eight metamorphic relations governing intra- and inter-triplet element diversity. Leveraging these relations, we can quantify the weaknesses of existing ASTE methods and large language models (LLMs) in capturing diversity. Furthermore, we devise a quantitative reflection strategy that flexibly incorporates linguistic constraints into both the training and reasoning phases. Extensive experiments on four ASTE datasets and three mental happiness datasets demonstrate that our MaRS outperforms several state-of-the-art models in F1-score and reduces violation rates by up to 10.80% for ASTE models and 17.29% for LLMs.