Neural Cellular Automaton Achieves 100% on 11 of 17 SLOG Categories
A novel method for structural generalization in semantic parsing employs a neural cellular automaton (NCA) featuring a discrete bottleneck, achieving a perfect type-exact match in 11 out of 17 categories on the SLOG benchmark. Unlike earlier techniques that depend on manually crafted algebraic rules (AM-Parser) or struggle with generalization (Transformer-based models), this approach derives all compositional rules from data through local iterations. It surpasses AM-Parser in three categories, where AM-Parser's scores range from 0% to 74%, and demonstrates a standard deviation of 0.2 across 10 seeds compared to AM-Parser's 4.3. An examination of 5,539 failure cases identifies two specific mechanisms: unique combinations of wh-extraction context with fewer verb types, and modifiers located on the subject side of verbs.
Key facts
- NCA with discrete bottleneck achieves 100% on 11 of 17 SLOG categories
- No hand-written compositional rules required
- Outperforms AM-Parser on three categories (0-74%)
- Standard deviation 0.2 across 10 seeds vs AM-Parser's 4.3
- All 5,539 failures due to two mechanisms
- Failures involve wh-extraction with reduced verb types
- Failures involve modifiers on subject side of verbs
- Published on arXiv:2604.26157
Entities
Institutions
- arXiv