PAST-TIDE introduces statement tuning with prototypical contrastive learning for stance detection

The authors introduce PAST-TIDE, a stance detection system designed for the StanceNakba Shared Task at NakbaNLP@LREC-COLING 2026. The core innovation is statement tuning, which redefines stance as cloze-style masked language modeling (MLM) to map label words via the pre-trained MLM head instead of using a randomly initialized classification head.

• PAST-TIDE employs prototypical contrastive learning with learnable class prototypes for batch-size independent training.
• The system utilizes topic-conditional layer normalization to handle cross-topic Arabic stance detection.
• It achieves macro-F1 scores of 0.75 for Subtask A and 0.74 for Subtask B on the official leaderboard.

The results indicate that minimal architectural additions to a pre-trained model can remain competitive in low-resource settings.