@inproceedings{he-zhou-2026-ynu,
    title = "{YNU}-{ABSA} at {S}em{E}val-2026 Task 3: A Unified Pipeline for Continuous and Structured Dimensional {ABSA}",
    author = "He, Qimao  and
      Zhou, Xiaobing",
    editor = "Kochmar, Ekaterina  and
      Ghosh, Debanjan  and
      North, Kai  and
      Komachi, Mamoru",
    booktitle = "Proceedings of the 20th {I}nternational {W}orkshop on {S}emantic {E}valuation (2026)",
    month = jul,
    year = "2026",
    address = "San Diego, California, USA",
    publisher = "Association for Computational Linguistics",
    url = "https://preview.aclanthology.org/ingest-acl-workshops/2026.semeval-1.150/",
    pages = "1100--1105",
    ISBN = "979-8-89176-414-9",
    abstract = "Dimensional Aspect-Based Sentiment Analysis (DimABSA) aims to jointly model continuous Valence{--}Arousal (VA) regression and structured sentiment extraction at the aspect level in multilingual settings, requiring both fine-grained emotion modeling and structural consistency. Prior approaches often separate regression and extraction or rely on stagewise pipelines, which may limit numerical stability and structural alignment. To address this challenge, we propose a unified pipeline for all three subtasks of DimABSA Track A.Although Task 1 and Task 2/3 use different backbone architectures, they are integrated through consistent preprocessing, a shared dimensional sentiment perspective, and unified post-processing principles. For Task 1, we enhance aspect{--}context interaction via aspect-conditioned cross-attention and attention pooling, together with bounded output mapping and lightweight calibration for stable VA prediction.For Task 2/3, we formulate triplet and quadruplet prediction as constrained conditional generation with LoRA fine-tuning and structural validation. Experiments show consistent improvements across languages, including lower RMSE, higher correlation, and better cF1. Error analysis further shows that Arousal remains more difficult than Valence."
}