@inproceedings{s-s-2026-pixel,
    title = "Pixel Phantoms at {S}em{E}val-2026 Task 3: Language-Specific Transformer Regression for Dimensional Aspect-Based Sentiment Analysis",
    author = "S, Jithu Morrison  and
      S, Abisha Rose",
    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-1.112/",
    pages = "803--810",
    ISBN = "979-8-89176-414-9",
    abstract = "Aspect-Based Sentiment Analysis (ABSA) has traditionally relied on discrete polarity labels (positive, negative, or neutral) which fail to capture the continuous, multidimensional nature of human emotion. SemEval-2026 Task 3, Dimensional Aspect-Based Sentiment Analysis (DimABSA), addresses this limitation by requiring systems to predict continuous Valence (pleasantness) and Arousal (intensity) scores on a 1{--}9 scale for specific aspect terms in text, across 15 language{--}domain combinations in two tracks. Prior approaches to multilingual ABSA have largely depended on single generic multilingual encoders applied uniformly across languages, ignoring language-specific linguistic structures. The Pixel Phantoms system takes a language-aware strategy, selecting dedicated language-specific pre-trained transformer models for each language, including {\textbackslash}url{\{}cl-tohoku/bert-base-japanese-v3{\}} for Japanese, {\textbackslash}url{\{}DeepPavlov/rubert-base-cased{\}} for Russian, {\textbackslash}url{\{}bert-base-chinese{\}} for Chinese, and a Davlan Swahili mBERT variant for Swahili, and falling back to {\textbackslash}url{\{}xlm-roberta-base{\}} for morphologically complex low-resource languages such as Tatar and Ukrainian. All models share a common regression architecture: a dual-pooling head combining CLS and mean-pooled representations, trained with a composite MSE + MAE loss and aspect-prompted input formatting. We participated in both Track A (10 combinations) and Track B (5 combinations), with our strongest result in Japanese Hotel (rank 13/21, RMSE 0.7297) and competitive performance in Chinese restaurant (RMSE 0.9823 vs. Baseline Kimi-K2 Thinking 1.8959). We also analyze failure modes in low-resource languages and domain-shifted settings, highlighting where multilingual transfer remains brittle. Overall, the results indicate that language-specific encoders deliver consistent gains over generic multilingual baselines in dimensional sentiment regression."
}