Reducing content bias in syllogistic reasoning via semantic abstraction. Experiments based on fine-tuning DeBERTa showed that abstraction-based augmentation reduces content effect and improves accuracy, leading to competitive performance on the official leaderboard.
