Reference Checker (Local App)

A local tool to verify references in a manuscript and flag potential mis-citations.

What it does

1. Extracts references from an uploaded PDF (or accepts a BibTeX file).
2. Verifies existence of each bibliographic entry via DOI / PMID / arXiv lookup (Crossref, PubMed, arXiv).
3. Checks claim support (beta): finds in‑text citation sentences, retrieves the cited source (abstract or open access text), and runs semantic matching + NLI to label as Supported / Not Found / Contradicted (risk).
4. Produces an interactive report plus CSV/JSON exports.

⚠️ Claim checking is best‑effort. It uses abstracts or OA text when available and may be limited by PDF parsing and access constraints.

Quickstart

python -m venv .venv
source .venv/bin/activate  # Windows: .venv\Scripts\activate
pip install -r requirements.txt
streamlit run app.py

Optional environment variables

Create a .env file in this folder to improve retrieval:

UNPAYWALL_EMAIL=your_email@example.com   # for Unpaywall OA lookups
NCBI_EMAIL=your_email@example.com        # for PubMed polite usage
CROSSREF_MAILTO=your_email@example.com   # for Crossref polite usage

How it works (high level)

• Parsing: finds the "References" section in the PDF; extracts DOIs/PMIDs/arXiv IDs; queries Crossref/OpenAlex/PubMed when missing.
• Existence check: resolves identifiers and basic metadata.
• Claim extraction: detects in‑text markers like [12], (Smith, 2020), Smith et al. (2020); grabs the surrounding sentence(s) as the claim.
• Support check: embeds claim and source sentences with Sentence Transformers and runs MNLI when available. Labels each claim–source pair and shows top matching snippets.
• Outputs: interactive table, per‑citation details, and downloadable CSV/JSON.

Notes

• PDF parsing is noisy; for best results, upload clean PDFs or add a BibTeX file of the references.
• Heuristics support common citation styles (numeric, author‑year). Edge cases may need manual review.
• The NLI model (roberta‑large‑mnli or bart‑large‑mnli) downloads on first run (~1.3GB). If unavailable, the app falls back to similarity‑only heuristics.