@inproceedings{chen-2026-pali,
    title = "{PALI} at {S}em{E}val-2026 Task 3: {L}o{RA} Fine-Tuning with Validation for {D}im{ABSA}",
    author = "Chen, Cheng",
    editor = "Kochmar, Ekaterina  and
      Ghosh, Debanjan  and
      North, Kai  and
      Komachi, Mamoru  and
      Zampieri, Marcos",
    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-brigap/2026.semeval-1.93/",
    doi = "10.18653/v1/2026.semeval-1.93",
    pages = "641--649",
    ISBN = "979-8-89176-414-9",
    abstract = "We describe the PALI system submitted to SemEval-2026 Task{\textasciitilde}3 (Dimensional Aspect-Based Sentiment Analysis), which requires predicting valence{--}arousal (VA) scores and extracting structured sentiment tuples across multiple languages.Our final system centers on LoRA fine-tuning of Qwen3-32B using Llama-Factory, together with data conversion/cleaning, multilingual data-mixing strategies, and inference-time validation and repair.We additionally explored retrieval-based few-shot prompting with BGE-M3, but found it less effective for learning consistent VA scoring preferences.On Track{\textasciitilde}A, our final system uses per-language LoRA adapters that mix all subtasks per language for a better trade-off between performance and efficiency.On the official test set, we achieve average per-language scores of 1.2071 RMSE{\textbackslash}VA for Subtask{\textasciitilde}1 and 0.5641/0.4905 cF1 for Subtask{\textasciitilde}2/3.On the development set, we find that per-language-per-task adapters further improve extraction cF1 but are less attractive in terms of training and deployment cost.For Track{\textasciitilde}B, we report results for VA prediction on five languages and two domains."
}