@inproceedings{satish-kumar-joshi-2026-ecoaffecttrack,
    title = "{E}co{A}ffect{T}rack at {S}em{E}val-2026 Task 2: A Hierarchical {D}e{BERT}a-Transformer Framework with {CCC} Optimization for Longitudinal Affect Modeling",
    author = "Satish Kumar, Diya  and
      Joshi, Om",
    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-1.77/",
    pages = "540--545",
    ISBN = "979-8-89176-414-9",
    abstract = "This submission proposes a hierarchical framework for longitudinal affect modeling, specifically designed for predicting variations in emotional valence and arousal over time. The system utilizes a DeBERTa-v3 encoder backbone optimized with a differentiable Concordance Correlation Coefficient (CCC) Loss for affect assessment (Subtask 1). This approach prioritizes capturing the ``shape'' and trend of emotional trajectories over absolute point-wise accuracy, yielding a significant performance gain over standard Mean Squared Error.For state change forecasting (Subtask 2A), the framework employs a Transformer-based temporal forecaster with positional encoding to account for inter-subject variability in emotional baselines. Disposition profiling (Subtask 2B) is addressed using a deep attention network that aggregates historical embeddings to identify emotionally informative essays. Experimental results from the official competition indicate that aligning the loss function with evaluation metrics and utilizing task-specific temporal modeling are essential for robust performance in longitudinal emotion recognition."
}