This New Research Could Fix AI's Confusion With New Word Combos
Why can't your phone understand 'blue apple'? Science is on it.
A new paper in computational linguistics asks a different question about compositional generalization: instead of measuring model accuracy, which structural or lexical identifications make held-out COGS examples admissible from the structures seen in training? The authors represent sentences as functors from syntactic addresses to lexical tokens, then use selective collapses to induce Kan extensions that propagate observed associations. Across 21 COGS generalization types, admissibility follows distinct identification profiles, while residual failures separate unsupported structural templates. These data-side diagnoses characterize what the training corpus licenses under specified identifications — without training a predictive model.
- AI models fail at new word combinations because of missing patterns in training data, not just weak algorithms.
- The researchers tested 21 puzzle types from the COGS benchmark and showed exactly which ones trip AI up.
- This math-based approach could help engineers fix training data first, saving time and making AI smarter without retraining.
- Category theory and 'Kan extensions' are advanced math, but the payoff is a clearer map of AI's limits.
Why It Matters
If AI can finally understand new word combinations, your virtual assistant will get your meaning more often — and waste less of your time.