Claude Models Outperform Specialized Systems in MRS-to-Text Generation
September 25, 2026
Claude Opus 5 achieves 76.3 BLEU on the English Resource Grammar's MRS-to-text generation task, surpassing a specialized system trained on 72k pairs which achieved 66.1 BLEU. The evaluation shows high-performing zero-shot capabilities for converting formal semantic graphs back into English sentences.
HOW THIS AFFECTS YOU
●
researcherThis suggests that large-scale pre-training may provide better linguistic mapping capabilities than task-specific sequence-to-sequence training.