On the Role of Directionality in Structural Generalization

The article introduces a redesigned symbolic backend for AM-Parser that utilizes CCG directed types to better handle directional distinctions in structural generalization tasks like modifier position shifts.

• The system uses deterministic CKY and a single linear decoder with 30K learnable parameters.
• It achieves 75.9±6.4% LF exact match, surpassing the previous AM-Parser score of 70.8±4.3%.
• Gains are highly directional, with the CCG system outperforming AM-Parser on all 5 position-shift categories by +29.9pp.
• Replacing the BERT-base encoder with DeBERTa-v3-large yields 90.7±4.9%, shifting the bottleneck to the neural layer.

This approach demonstrates that directional representations improve structural generalization, particularly for position-shift tasks, while complementary gains from larger encoders address recursive-depth categories.