Most Extracted Research Terms Describe One Paper
Published · paper counts are current, not as-published
The keywords on this site are not chosen by authors. They are pulled out of each paper's own text by machine, and this corpus holds 214,531 of them across 224,410 papers. Those two numbers being close is the first thing to explain: a vocabulary that grows at roughly the rate the corpus does is not being shared between papers.
Read that as a fact about research and you would be reading past the instrument. It is a fact about what this extractor produced. The question worth asking is whether the tail is an artefact of how the extraction works, and that is testable rather than arguable - so below are the three ways it could be an artefact, each one measured.
One: It Could Be Spelling
The obvious objection is that a tail like this is the same idea written several ways: plurals, hyphens, British and American spellings. This corpus merges those variants on every import, and what survives is 0 singular/plural pairs still carrying papers on both sides. The tail is not typos.
Two: It Could Be Sliced Phrases
The stronger objection is that the extractor cuts one phrase into several, so a term that looks unique is really a piece of a term used often - "effect estimation" is not a rare concept if it only ever occurs inside "causal effect estimation". That is checkable: take every multi-word term used once, and ask whether it appears verbatim inside a longer term that is itself used more than once.
Of 144,002 such terms, 2,060 do - 1.4%. Slicing is real and it is marginal. It does not account for the tail.
Three: It Could Be Phrase Length
Longer phrases are rarer by arithmetic, whatever they mean, so a tail made mostly of four-word strings would say nothing about research. If that were the mechanism, the share used once would climb with every word added.
| Words in term | Terms | Used once or never |
|---|---|---|
| 1 | 3,923 | 72.6% |
| 2 | 150,361 | 64.8% |
| 3 | 55,268 | 77.2% |
| 4 or more | 4,979 | 78.3% |
If length were the mechanism, the shortest terms would be the reused ones. They are not: a one-word term describes a single paper 72.6% of the time, within a few points of the longest group in the table. The share moves with length, but never far enough, and never in the one direction a length explanation needs. Whatever produces this tail, it is not how many words the terms have.
What That Does to a List of Top Terms
The keyword cloud on the keywords page shows the 200 most used terms. That is 0.09% of the vocabulary - and it accounts for 27% of all keyword usage in the corpus.
Both halves of that sentence matter. A list of top terms is not a summary of what gets written about, because it omits 99.91% of the vocabulary. But it is not arbitrary either, because that sliver of the words carries more than a quarter of the usage. The head is steep: large language model (14,686), neural network (7,070), representation learning (6,926).
What This Still Does Not Show
- Three artefacts ruled out is not all of them. Variant merging catches plural and punctuation forms; it does not catch synonyms, acronyms or renamings, so some of the tail is still one idea under two names.
- A term used once is not a concept used once. It may mean one paper coined it, or that only one paper phrased it that way - this measures phrasing, and phrasing is what an extractor can see.
- Not a controlled vocabulary. For the curated side, the topic hierarchy assigns every paper to a small fixed set instead - a different instrument answering a different question.
The numbers here are read from the database when you load the page.