README.md

---
license: cc-by-4.0
language: en
task_categories:
  - text-classification
  - question-answering
  - text-retrieval
---


# DefExtra

<p align="center">
  <a href="https://arxiv.org/abs/2602.05413"><img src="https://img.shields.io/badge/arXiv-2602.05413-b31b1b" alt="arXiv:2602.05413"></a>
  <a href="https://sigir.org/"><img src="https://img.shields.io/badge/SIGIR%202026-under%20review-0054a6" alt="SIGIR 2026 under review"></a>
  <a href="https://huggingface.co/datasets/mediabiasgroup/DefExtra"><img src="https://img.shields.io/badge/HF%20Dataset-DefExtra-ff9d00" alt="HF Dataset DefExtra"></a>
  <a href="https://huggingface.co/datasets/mediabiasgroup/DefSim"><img src="https://img.shields.io/badge/HF%20Dataset-DefSim-ff9d00" alt="HF Dataset DefSim"></a>
  <a href="https://media-bias-group.github.io/SciDef-ProjectPage/"><img src="https://img.shields.io/badge/Project%20Page-SciDef-2e7d32" alt="SciDef Project Page"></a>
  <a href="https://doi.org/10.5281/zenodo.18501198"><img src="https://img.shields.io/badge/Zenodo-10.5281%2Fzenodo.18501198-1682D4?logo=zenodo" alt="Zenodo DOI: 10.5281/zenodo.18501198"></a>
  <a href="https://github.com/Media-Bias-Group/SciDef"><img src="https://img.shields.io/badge/Code-GitHub-181717?logo=github" alt="Code on GitHub"></a>
</p>

## Overview

DefExtra contains 268 definition records (term, definition, context, type) from 75 papers. **We do not ship excerpts from papers** due to copyright. Instead, we ship markers and scripts that let users hydrate the dataset from their own PDFs.

Why this workflow:

- We cannot redistribute copyrighted excerpts.
- We therefore ship **only localization markers** plus scripts to reconstruct the text from user‑supplied PDFs.

## Examples (from our own papers; after hydration)

| Source | Concept | Definition | Context (excerpt) |
| --- | --- | --- | --- |
| [https://aclanthology.org/2024.lrec-main.952](https://aclanthology.org/2024.lrec-main.952) | `media bias` | “a skewed portrayal of information favoring certain group interests, which manifests in multiple facets, including political, gender, racial, and linguistic biases.” | “Media bias is a skewed portrayal of information favoring certain group interests … Such subtypes of bias … make the classification of media bias a challenging task.” |
| [https://arxiv.org/abs/2312.16148](https://arxiv.org/abs/2312.16148) | `spin bias` | “a form of bias introduced either by leaving out necessary information or by adding unnecessary information.” | “Spin Bias describes a form of bias introduced either by leaving out necessary information … or by adding unnecessary information.” |

## Quickstart (DefExtra hydration)

1) Put PDFs in `pdfs/` (filename should match `paper_id`, DOI/PII alias, or arXiv ID).
2) Start a GROBID server (see `docs/defextra_hydration.md`).
3) Hydrate:

```bash
uv run python scripts/hydrate_defextra.py \
  --legal-csv data/defextra_legal.csv \
  --pdf-dir pdfs \
  --grobid-out grobid_out \
  --output-csv defextra_hydrated.csv \
  --report defextra_hydrated_report.txt \
  --require-complete
```

## Getting PDFs

- See `docs/get_pdfs.md` for sources and a helper script that lists required PDFs.
- `defextra_required_pdfs.csv` and `defextra_required_pdfs.md` are precomputed lists.

## Environment (uv)

- This repo ships a `pyproject.toml` with all dependencies.
- Run any script with `uv run python ...` and uv will resolve/install deps.

## Data files

- `data/defextra_legal.csv` / `data/defextra_legal.parquet`: DefExtra markers (no excerpts).

## Hydrated columns

The hydrated output (e.g., `defextra_hydrated.csv`) matches the schema below.
Full legal marker columns are documented in `docs/defextra_hydration.md`.

| Column | Description |
| --- | --- |
| `paper_id` | Paper identifier (often a Semantic Scholar ID, DOI, or arXiv ID). |
| `paper_title` | Paper title. |
| `paper_doi` | DOI (if available). |
| `paper_arxiv` | arXiv ID or URL (if available). |
| **`concept`** | Term / concept being defined. |
| **`definition`** | Definition text (hydrated from PDFs). |
| **`context`** | Context excerpt (hydrated from PDFs). |
| **`definition_type`** | Definition type (e.g., explicit / implicit). |
| `source_file` | Source JSON filename used during curation. |
| **`is_out_of_domain`** | Boolean flag for out‑of‑domain papers. |

## Scripts

- `scripts/hydrate_defextra.py`: hydrate DefExtra from PDFs + GROBID.
- `scripts/pdf_to_grobid.py`: batch GROBID runner (requires a running GROBID server).
- `scripts/list_defextra_pdfs.py`: list required PDFs + download links.
- `scripts/build_defextra_test_pdfs.py`: build a test PDF set from a larger PDF pool.
- `scripts/report_defextra_status.py`: summarize missing items by paper/definition.

## Documentation

- [`docs/defextra_hydration.md`](docs/defextra_hydration.md) (technical details, CLI flags, markers).
- [`docs/get_pdfs.md`](docs/get_pdfs.md) (how to find PDFs).
- [`docs/mismatch_examples.md`](docs/mismatch_examples.md) (mismatch types with short excerpts).

## Expected minor mismatches

- Small differences vs. the manual reference can occur due to PDF/GROBID text normalization.
- Typical cases: line‑break hyphenation, spacing around numbers, citation formatting.
- These are documented and do not affect the ability to hydrate all entries.

## Notes

- Hash IDs are typically Semantic Scholar paper IDs; many PDFs can be obtained from Semantic Scholar.
- If you see PDF hash mismatch warnings, verify you have the correct paper version and rerun with `--allow-pdf-hash-mismatch` only after manual inspection.
- The script was largely produced using LLMs for robustness.

## Citation

```bibtex
@misc{kucera2026scidefautomatingdefinitionextraction,
      title={SciDef: Automating Definition Extraction from Academic Literature with Large Language Models}, 
      author={Filip Ku\v{c}era and Christoph Mandl and Isao Echizen and Radu Timofte and Timo Spinde},
      year={2026},
      eprint={2602.05413},
      archivePrefix={arXiv},
      primaryClass={cs.IR},
      url={https://arxiv.org/abs/2602.05413}, 
}
```