HassanB4/dialectsenteval2026-arabart-prefix-tag

Fine-tuned on DialectSentEval 2026 Subtask 2 (Arabic Dialect Sentiment Swap): given an Arabic sentence and its polarity, rewrite it to express the opposite polarity while preserving meaning. One of 6 trained backbones behind NAMAA Community's submission (formerly S05_arabart_prefix_tag in internal experiment logs).

• Base model: moussaKam/AraBART
• Input format: prefix_tag
• Dev Sentiment Style Accuracy: 76.87
• Task: DialectSentEval 2026, Subtask 2 — https://ezzini.github.io/DialectSentEval2026/
• Code: https://github.com/HasanBGit/DialectSentEval-Arabic-Sentiment-Swap

This checkpoint alone is not our best system — reranked sampling and cross-backbone ensembling on top of these checkpoints reach 96.78 development accuracy and 0.7597 on the official hidden test set. See the repository for the full pipeline, results, and the honest dev/test generalization-gap analysis.

Citation

@inproceedings{namaa2026dialectsenteval,
  title = {NAMAA at DialectSentEval 2026: From Reranked Sampling to Cross-Backbone Ensembling for Arabic Dialectal Sentiment Swap},
  author = {TODO Author List},
  booktitle = {Proceedings of the Fourth Arabic Natural Language Processing Conference},
  address = {Budapest, Hungary},
  publisher = {Association for Computational Linguistics},
  year = {2026},
}