HIPE 2026 Relation Extraction Pipeline

Clean source repository for the INSA Lyon HIPE 2026 person-place relation system. The repository is centered on one main Python package, a small set of runnable scripts, and the benchmark assets that are directly needed to reproduce the pipeline.

Repository layout

hipe/                 core package
  data/               pair-level loaders and split utilities
  preprocessing/      dataset enrichment and feature preparation
  features/           temporal and handcrafted feature helpers
  mask/               hmBERT MASK-template representations
  llm/                prompting, parsing, and LLM inference helpers
  retriever/          retrieval and embedding utilities
  stacker/            final decision-layer logic
  submission/         official submission writing and validation
scripts/              runnable experiments and export utilities
tests/                unit tests for core data and submission logic
data/                 benchmark inputs, derived data, and split files
docs/                 concise technical notes kept with the source

Data layout

• data/benchmark/train/: official multilingual HIPE 2026 training inputs
• data/benchmark/official_test/: official evaluation inputs, including the French surprise split
• data/derived/dataset_reference.jsonl: enriched pair-level dataset used by the main pipeline
• data/splits/v1_baseline_train_test_ids.csv: local development split used by several scripts
• data/reference/: released gold files used for post-hoc scoring
• data/schemas/: official JSON schema
• data/examples/: small supporting examples

See data/README.md for a short description of each subfolder.

Main entry points

• scripts/run_full_pipeline.py: end-to-end prediction workflow
• scripts/make_submission.py: convert pair-level predictions into official submission JSONL files
• scripts/fetch_data.py: download benchmark files expected by the pipeline
• tests/test_submission.py: core checks for official-output formatting

Notes

• docs/notes/ keeps project notes that are still useful for understanding the final system.
• Large generated artefacts, caches, notebooks, intermediate outputs, and local backup folders are intentionally excluded from this cleaned repository.