Structured Example Selection for Few-Shot Relation Extraction

Code and data for:

Aunabil Chakma, Mihai Surdeanu, and Eduardo Blanco. 2026.

Structured Semantic Information Helps Retrieve Better Examples for In-Context Learning Applied to Few-Shot Relation Extraction.

ACL 2026.

This repository contains the core implementation, prompts, data, and rules used in the paper.

Setup

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

If a Hugging Face model requires authentication, set HF_TOKEN in your environment. Do not place tokens in source files.

Select retrieved examples

select_examples.py accepts a file containing candidate examples, scores, and embeddings:

python select_examples.py \
  --input external_candidates.jsonl.gz \
  --output selected.jsonl.gz \
  --strategy kmeanspp_farthest \
  --count 4 \
  --threshold 0.6

Use --count 4 for 5-shot experiments and --count 9 for 10-shot experiments.

Run inference

python run_inference.py \
  --episodes <episodes.jsonl.gz> \
  --supports <support_examples.jsonl.gz> \
  --queries <query_examples.jsonl.gz> \
  --relations <relations.json> \
  --model Qwen/Qwen3-4B \
  --entity-filter \
  --output outputs/predictions.jsonl.gz

The same command supports FewRel and authorized TACRED data using the common JSONL schema. Pass selected or generated examples through --extra-examples for 5- or 10-shot experiments.

Evaluate

python evaluate.py outputs/predictions.jsonl.gz

Citation

If you use our work, please cite our paper using the BibTeX below:

@inproceedings{chakma-etal-2026-structured,
    title = "Structured Semantic Information Helps Retrieve Better Examples for In-Context Learning Applied to Few-Shot Relation Extraction",
    author = "Chakma, Aunabil  and
      Surdeanu, Mihai  and
      Blanco, Eduardo",
    editor = "Liakata, Maria  and
      Moreira, Viviane P.  and
      Zhang, Jiajun  and
      Jurgens, David",
    booktitle = "Proceedings of the 64th Annual Meeting of the {A}ssociation for {C}omputational {L}inguistics (Volume 1: Long Papers)",
    month = jul,
    year = "2026",
    address = "San Diego, California, United States",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2026.acl-long.1664/",
    doi = "10.18653/v1/2026.acl-long.1664",
    pages = "35947--35971",
    ISBN = "979-8-89176-390-6"
}