ML Anthology

An anthology for machine learning research.

What is this?

ML Anthology indexes papers from ML conferences and journals into a single browsable, searchable site with consistent metadata and human-readable citation keys.

It exists because ML proceedings are fragmented across a dozen different sites with inconsistent interfaces, no unified search, and no persistent identifiers. NLP solved this years ago with ACL Anthology. This is the equivalent effort for machine learning.

Citing an ML paper currently usually means something like getting 10.5555/3295222.3295349 (Vaswani et al.), NEURIPS2018_5a4be1fa (this is the NTK paper, but you'd never know), pmlr-v151-truong22a, or 7780459 (Resnet, believe it or not) — equally opaque strings, none of which tell you what paper you're looking at. For NeurIPS you have to download a .bib file from each one, open it, and hope you remember which download was which. It sucks.

Features

• Human-readable BibTeX keys — vaswani2017neurips-attention instead of opaque hashes or integers
• Deterministic URLs — /venue/year/key, computable from paper metadata alone
• Citation export — BibTeX, plain text, and Markdown with copy buttons
• Full-text search — via Pagefind, entirely client-side
• Author index — many authors with per-author publication pages
• Dark and light mode — follows system preference, looks nice
• Static site — fast, no backend, trivially cacheable

URL scheme

mlanthology.org/icml/                                     → latest ICML proceedings
mlanthology.org/icml/2017/                                → ICML 2017 paper listing
mlanthology.org/icml/2017/arora2017icml-generalization    → specific paper at ICML 2017
mlanthology.org/authors/                                  → author index
mlanthology.org/authors/a/                                → authors by last name starting with a
mlanthology.org/authors/a/arora-sanjeev/                  → author page for sanjeev arora

Every prefix is a valid page, you can feasibly construct the permalink by guessing if you know the title, year, author, and venue.

ResNet, for example, is he2016cvpr-deep because it appeared at CVPR in 2016, the first author was Kaiming He, and the first word of the title was "Deep".

BibTeX key format

Keys follow the pattern:

{lastname}{year}{venue}-{titlekeyword}

• lastname — first author's lowercased surname
• year — four-digit publication year
• venue — lowercased venue abbreviation
• titlekeyword — first substantive word from the title (stopwords filtered)

Collisions are resolved with a short suffix. The generation is fully deterministic from the paper metadata.

API Access?

I didn't make one, but every paper page embeds structured metadata as JSON-LD in the document head.

Notice that:

import requests, json, re

def get_paper(key):
    r = requests.get(f"https://mlanthology.org/{key}")
    m = re.search(r'<script type="application/ld\+json">(.*?)</script>', r.text, re.DOTALL)
    return json.loads(m.group(1))

>>> get_paper("neurips/2017/vaswani2017neurips-attention")
{
  "@type": "ScholarlyArticle",
  "headline": "...",
  "author": [...],
  "datePublished": "2017",
  "description": "...",
  ...
}

Data sources

ML Anthology is built on top of open data from:

• DBLP, Semantic Scholar, and OpenAlex — bibliographic metadata
• OpenReview — ICLR and TMLR metadata
• NeurIPS — NeurIPS metadata
• PMLR — ICML, AISTATS, COLT, and other proceedings
• CVF Open Access — computer vision proceedings

License

Apache 2.0. See LICENSE.

Paper metadata is aggregated from the sources above under their respective licenses. ML Anthology does not host PDFs (yet) and papers link to their original sources.