Few-Shot Relation Extraction with AllenNLP

The repo contains additional functionality, not implemented in AllenNLP, for few-shot relation extraction / relation classification.

• Dataset reader for FewRel dataset
• N-way K-shot iterator, creating batches according to the few-shot scenario
• Few-shot relation classifier, configurable via jsonnet files (see experiments/config/ for examples)
• Relative offset embeddings (required for non-sequential models, i.e. CNN)
• Additional models for few-shot classification
  • Prototypicial Network

Setup

$ pip install -r requirements.txt

Run experiment

$ allennlp train --include-package fewrel experiments/configs/baseline.jsonnet -s <RESULT DIR>

References

• Han, Xu et al. “FewRel: A Large-Scale Supervised Few-Shot Relation Classification Dataset with State-of-the-Art Evaluation.” EMNLP (2018).
• Snell, Jake et al. “Prototypical Networks for Few-shot Learning.” NIPS (2017).