@inproceedings{sukhodolsky-etal-2026-bertkittens,
    title = "{B}ert{K}ittens at {S}em{E}val-2026 Task 3: Multi-Domain Aspect Sentiment with {BERT}/{D}e{BERT}a Ensembles for {VA} Regression and Aspect{--}Opinion{--}{VA} Triplets",
    author = "Sukhodolsky, Arseny  and
      Salimgareev, Ruslan  and
      Ianshina, Tatiana",
    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-1.380/",
    pages = "3027--3034",
    ISBN = "979-8-89176-414-9",
    abstract = "Our system is built on transformer encoders (BERT and DeBERTa) fine-tuned in a multi-task learning framework. For the regression subtask (evaluated with RMSE), we jointly predict Valence{--}Arousal (VA) scores and token-level opinion spans using a shared encoder with task-specific output heads. This formulation introduces auxiliary supervision at the token level, which stabilizes training and improves regression accuracy compared to single-task optimization.When gold abstracts and opinion annotations are provided, our models achieve strong performance. However, in fully end-to-end settings requiring automatic span extraction, performance degrades substantially due to error propagation from token-level predictions.Our findings highlight the benefits of joint affective regression and span modeling, while exposing the limitations of transformer-based sequence labeling under strict end-to-end evaluation constraints."
}