zjunlp/AdaKGC

Combines T5 encoder-decoder architecture with UIE-style tokenization and learnable soft prompts to enable unified handling of entity, relation, and event extraction across diverse schemas. Supports schema adaptation through iterative dataset construction modes (H/V/M/R) and negative sampling, with evaluation metrics spanning spot-F1 (entity/trigger detection) and asoc-F1 (argument/role linking). Integrates with HuggingFace T5 models and provides both single-dataset and multi-iteration inference pipelines on benchmarks like ACE05, Few-NERD, and NYT.

No commits in the last 6 months.