scholarmend

Recover true bibliographic metadata for Google Scholar exports, so that screening tools reason over facts rather than over Scholar's truncations.

Usage

pip install scholarmend
scholarmend --input path/to/scholar-exports --out out

--input takes a directory of RIS files exported from Google Scholar, or a single .ris. For development, clone the repo and pip install -e ".[dev]".

Tier 1 needs no configuration and resolves venue, year and track for 77% of a Scholar corpus. Tier 2 needs an OpenReview account:

export SCHOLARMEND_OPENREVIEW_USER='you@example.edu'
export SCHOLARMEND_OPENREVIEW_PASSWORD='...'

Semantic Scholar, the last resort for records nothing else resolves, works without a key but shares a slow anonymous pool that often answers 429. A free API key gives you your own 1 request per second:

export SCHOLARMEND_S2_KEY='...'

Every response is written to --cache (default .scholarmend-cache). Commit it: a rerun then reproduces byte-identically and makes no API calls.

scholarmend --input path/to/scholar-exports --out out --offline

--offline never reaches the network. A record whose lookup is missing from the cache falls back to tier 1, and the run exits 1 to say it was partial.

Outputs

The problem

Google Scholar's .ris export is the standard input to a systematic review, and it is lossy in ways that stay invisible until they corrupt a result. Measured on a 2,413-record corpus of NeurIPS, ICLR and ICML search results:

One record shows the whole problem, and its own answer:

JF  - … Neural Information …
AU  - ...
PY  - 2026///
UR  - https://proceedings.neurips.cc/paper_files/paper/2025/hash/4da4f3c0…-Abstract-Datasets_and_Benchmarks_Track.html

Scholar reports an ellipsis for the venue, the wrong year, and no track. The URL in the same record states NeurIPS, 2025, Datasets and Benchmarks Track.

The approach

The identifier in this corpus is not a DOI — there are two — it is the URL path. Offline URL mining yields a key for 99% of records and fully resolves venue, year and track for 77%, with no network access at all. Records are then escalated per field through cached, credentialed lookups only where mining leaves a gap.

Every claim is retained with its source, tier, confidence and evidence, so the losing values stay auditable rather than being overwritten silently.

Abstracts, on request. Scholar's AB is a search snippet on every record — fragments joined by … around the query terms — and that is what a screening tool shows reviewers. --abstracts replaces it with the paper's abstract, from the proceedings page, the OpenReview submission note, or Semantic Scholar, in that order, admitting each only when the source names the record's own title. Run it on the deduplicated, triaged upload rather than the whole corpus:

scholarmend --input ../Trust-Evals-LitReview/out/clean.ris --out out-covidence --abstracts

report.txt lists every record whose abstract is still Scholar's snippet.

What it does not recover: authors. Scholar truncates 71% of author lists, and scholarmend leaves them as it found them: rewriting AU would restructure a repeated field, which the projection never does. See BACKLOG.md §5.

Querying a structured database instead does not work here, and the spec records the measurement: on a 30-title sample, OpenAlex matched 63% of titles and returned a correct conference venue for none of them, describing the arXiv preprint instead. The preprint carries a DOI and the proceedings version does not, so a DOI-anchored index indexes the preprint.

Validation

scholarmend is tested against hand-verified ground truth rather than against its own output. The Trust-Evals-LitReview review produced labels for 112 records that escaped an automated rule table and were resolved individually, with the evidence for each recorded.

What "settled" does and does not mean

The 104 figure counts records for which the miners extract something a live resolver consumes — a venue directly, or an OpenReview forum id, a PMLR volume or a PMC id. That is a claim about reach, not about correctness, and the two halves of it are verified to different depths:

• 90 records reached through an OpenReview venueid: venue, year and a workshop verdict, every one agreeing with the reviewers' hand-verified label. Zero disagreements, compared record by record.
• 10 PMLR volumes, all out of scope for this review — the Canadian Conference on AI, a workshop at ACML, an AAAI bridge programme, a parsimony conference. The resolver retrieves each proceedings title and deliberately declines to name a venue, rather than coerce an unfamiliar conference onto a known one. Verified against the volume the reviewers recorded in venue_true.
• 3 records via the PMC bridge, reproducing the reviewers' override decisions exactly: PMC volume 267 → PMLR v267 → ICML 2025, named; v287 (CHIL) and v297 (ML4H) retrieved and declined.
• 1 record reached by a sibling proceedings URL, not separately asserted.
• 8 records have no automated route at all: NSF landing pages, Google Books chapters, an SPIE paper, a PDF on a personal page. These needed human judgement before and still do — and the suite asserts they resolve to nothing, so the day a miner starts covering one, a test says so.

So the honest summary is that scholarmend determines and verifies the venue for 103 of the 112, reaches 1 more without a separate assertion, and leaves 8 untouched.

Run them with the review repository checked out alongside this one:

pytest tests/test_acceptance.py -v

They skip cleanly when it is not.

Relationship to other tools

• venuetriage — consumes scholarmend's output to separate workshop from main-track papers before Covidence.
• refaudit — verifies a finished bibliography against Crossref, OpenAlex and arXiv. Different job, different input; its DOI-centric resolvers reach only ~8% coverage on this corpus.

Licence

MIT.