What Counts as a Research Gap? Our Operational Definition

Our Research Gap Finder holds 120,000+ open gaps mined from 100,000+ papers. A question comes up often: what exactly counts as a "research gap" in that corpus, and what gets filtered out? This post gives the full operational definition, from the regex patterns that extract candidate sentences to the quality criteria that decide which ones reach you.

Why the definition matters

The phrase "research gap" is broad enough to mean almost anything: an unexplored question, a methodological limitation, a missing dataset, a population that has not been studied. That breadth is a feature when you are exploring a new area. It becomes a liability when a tool presents a sentence that says nothing more than "future work should be done" and calls it a gap.

Every corpus is shaped by the choices made at extraction. Ours is explicit. The rules are in code, and the code has been revised each time a real user saw something wrong. What follows is an account of that definition as it stands today.

Step 1: The linguistic signal

Every paper in the local library is processed at ingest by a Python extractor (library/scripts/gap_extractor.py). The extractor targets sentences in four section types: future_work, limitations, inline_gaps, and abstract. It applies regex patterns designed to match the linguistic shape of an open research question.

When the local library holds fewer than five candidates for a search query, journal/src/lib/research-gaps/external-gap-mine.ts applies the same patterns to paper abstracts retrieved in real time from an external bibliometric index, at zero LLM cost. These externally-mined gaps are badged amber in the results so you can distinguish them from the locally-harvested ones.

The patterns look for:

• Explicit future-work framing. "Future work should investigate", "remains to be explored", "warrants further study."
• Open-question language. "An open question", "an unresolved challenge", "poorly understood."
• Knowledge-absence markers. "No prior study has", "little is known about", "to the best of our knowledge."
• Narrative signals added in 2026. "A key challenge", "under-explored", "it remains unclear whether."

A sentence that matches any of these patterns is a candidate gap. It has not yet passed the quality filters.

Step 2: Seven classes of non-gap

Candidate sentences go through a pipeline of filters (journal/src/lib/research-gaps/sanitize-gap-text.ts). Each filter removes a class of false positive documented in a real user report.

Boilerplate. "Further research is needed in this area" states nothing. The extractor marks rows with boilerplate = 1; every downstream query requires boilerplate = 0 so these rows are never served.

Recommendations. A sentence from a paper's Recommendations section — "programs should be redesigned as continuous, structured processes" — is advice to practitioners, not a gap in the literature. These are dropped unless the sentence also contains an explicit absence marker ("no study has examined", "remains unknown"), in which case it is kept.

Meta-descriptions. "Finally, the article points out where future research in the field is needed" describes the paper's structure, not the science. Before this filter was added, nine of 68 externally-mined gaps served to real users (13.2%) were this class.

Contribution statements. "In this paper, we introduce an AI-driven framework that advances the ARHG model" describes what the paper did, not what the field lacks. The filter is conservative: a sentence that opens with contribution framing but goes on to state an absence ("however, no study has yet examined X") is kept.

Reference-list entries. The PDF extractor sometimes reads past a paper's last section into its references. Two independent citation signals are required before a candidate is dropped, so a genuine gap sentence that merely cites a paper is never removed.

Discourse fragments. A sentence lifted from mid-paragraph often begins with "However," or "Accordingly," — context-dependent, intelligible only alongside the sentences around it. These are dropped unless they state a concrete absence.

Extraction debris. CID-font failures, author contact blocks that bled into the extractor's reading window, all-caps journal-name runs — junk that is never a gap.

After filtering, roughly 120,000 open gaps remain: the library_gap table rows with still_open = 1 and boilerplate = 0.

Step 3: Four quality criteria

The Gap Finder's Pro tier (journal/src/lib/research-gaps/gap-criteria-packs.ts) exposes screening criteria that researchers can apply to their own search results. Four criteria are shared across every discipline:

Recency (2020 or later). A gap from a 2014 paper may have been addressed since. Source papers published 2020 or later are more likely to name open questions that are still open.

Well-scoped. "More research is needed in machine learning" is not a gap. A well-scoped gap is narrow enough to be addressable in one to three papers, with a concrete method and a defined evaluation.

Well-grounded. The gap should sit in an established research area, not a fringe one-off remark. Grounding is assessed from corroboration: a synthesis drawn from multiple source papers, or a gap whose topic matches a substantial body of library literature, passes. A gap that is the only mention of its topic across 4.5 million papers warrants caution.

Actionable. Can this gap be addressed by a tractable research project, with a defined dataset, a clear method, and a realistic timeline within one to two years?

Discipline-specific criteria extend these four. Engineering and physics add "simulation-friendly" and "benchmark available." Medicine adds "RCT-feasible" and "LMIC-applicable." Computer science adds "iso-compute comparable" and "reproducible code and data."

Step 4: From 120,000 gaps to 1,100+ citable pages

The 120,000+ individually mined gaps represent one sentence each. The 1,100+ gaps at /research-gaps/ are different: they are LLM-synthesised summaries drawn from multiple source papers, verified as grounded by the local library, and long enough to carry a real finding.

Each citable page names its source DOIs. You can trace the claim back to the paper it came from and, from there, to the journal, the methods section, and the data. That is the standard we hold ourselves to: a gap with no source is an assertion; a gap with a source is a starting point for a literature review.

The search at /research-gaps is free. The full synthesis table, which requires a generation pass, is metered. The gap pages themselves remain open and citable under CC BY 4.0.

What this means for authors

If you are looking for a direction to take a paper, or trying to position a manuscript against the existing literature, the corpus described here is a starting point. Search your topic free: the results table shows which gaps are grounded in the local library of 4.5 million papers, and which are thinner. Our Pre-Check tool can score a draft against the same evidence base before you send it anywhere.

If you spot a gap page that cites a result now addressed by newer work, the right action is to note the paper that addressed it: not to erase the question, but to sharpen the framing for the next one. Open science works that way.

Frequently asked questions

What is the library the gaps are mined from?

The local library holds 4.5 million papers, harvested from institutional sources and open-access repositories. The gaps in our corpus are mined from this library, not from external bibliometric indices used for citation counts and field statistics.

Why does the gap search sometimes return fewer results than expected?

When the local library holds fewer than five candidates for a query, the system falls back to real-time extraction from external paper abstracts. These externally-mined gaps are badged amber so you can distinguish them from the locally-harvested ones.

Can a gap be a limitation rather than a future-work statement?

Yes. Limitations and future-work sentences are treated as equivalent at extraction time. Both name something the paper did not or could not do, and both are potential starting points for new research.

How do I know a gap page has not already been answered?

The still_open flag is set at ingest and is not automatically updated against new literature. The recency filter (2020 or later) reduces the risk, but the most reliable check is your own search. The gap page gives you the source DOI and year; a forward-citation search in any database will show follow-up work. Our how to find a research gap guide covers that workflow in detail.

Are the gap pages available for citation?

Yes. The 1,100+ synthesis pages under /research-gaps/ are published under CC BY 4.0. Each page has a stable URL and a listed source DOI. A bulk export of the underlying library_gap table is not currently offered.

AI Review reports are written by AI agents. They are pre-submission feedback, not an editorial decision.