Analyzer · research papers
Claim–evidence map for a research paper
Paste a paper and see its main claims, the sentences linked to each one, how strongly each claim is worded and which claims have no linked sentence in the text. It maps wording and structure. It cannot tell you whether a claim is true.
Try it
Everything runs in this tab. The tool does not upload what you enter and stores nothing unless you export it.
Paper text
0 characters · limit 400,000At a glance
text patterns onlyClaims
wording and linked evidence, not truthPaper with anchors
select a claim or a sentence to highlight itEach sentence has an anchor, such as §Results ¶2 s1 · p. 3. Sentences that are linked to a claim have a dotted underline. Selecting a claim outlines it and underlines its evidence in blue.
Extractive summary
sentences from the paperStatistics notes
wording patterns, each with its sourceQuestions to ask the authors
generated from gaps in this textExport
files are made in this tabThe Markdown and JSON exports repeat the limit: they map wording and structure and do not say whether a claim is true. The CSV and the anchor list hold the claim data only.
How to use Claim–Evidence Map
- Paste or load the paper. Paste plain text, open a .txt or .md file, or load a text-based PDF of up to 80 pages.
- Press Analyze. A progress bar shows the work, and Cancel stops it. The tool does not save your text.
- Check each claim. Read its wording, its linked evidence and its anchor, and use Show in paper to see it in context. An unsupported marker means no linked sentence in this text, not a false claim.
- Export what you need. Copy the Markdown table, download the CSV or JSON, or copy the list of anchors.
What the map measures
The map reads one paper and reports four things about each claim sentence. It shows how strongly the sentence is worded. It shows which other sentences in the same text share its key terms. It rates how strong those sentences are as evidence. And it flags wording that is stronger than the evidence in the text allows.
Wording is matched against four word lists. The lists are this tool's own and overlap with the examples that writing guides give. The University of North Carolina Writing Center lists uncertainty words such as appears, seems, suggests and indicates. It also pairs absolute words with qualified ones, such as all with many, most or some (UNC Writing Center, Qualifiers).
| Level | Examples of words matched (and their forms) | What the level means here |
|---|---|---|
| hedged | may, might, could, possibly, likely, appears, seems, suggests, tends to | The writer marks uncertainty. |
| moderate | indicates, supports, associated with, is consistent with | The writer reports a pattern without claiming proof. |
| strong | demonstrates, shows that, establishes, proves, confirms | The writer claims the result is established. |
| absolute | always, never, all, none, every, undeniably | The writer claims the result holds without exception. |
The map does not decide whether a claim is true. It does not read the references, recompute a statistic, or know what the study measured. Everything it reports is a pattern in the words you pasted. For the reading method this supports, see How to read a research paper with a claim–evidence map.
How the scoring works
A sentence becomes a claim candidate when it has at least one cue: a contribution phrase such as "we show" or "our results indicate", a finding word such as "improved", "higher" or "state-of-the-art", or a comparison word together with a number. A candidate also needs a total score of 2 or more. Figure and table captions are never claims, though they can serve as evidence. Sentences of fewer than five words are not claims. Text before the first heading is scanned under the label "Start of text", so it can hold claims. Nothing after the References heading is scanned, and the Acknowledgements section is skipped.
| Signal | Points |
|---|---|
| Contribution phrase: we show, we find, we propose, our results indicate | +2 |
| Finding or comparison word: improve, reduce, higher, lower, outperform, novel, first | +1 |
| Comparison word together with a number: than, versus, more, fewer | +1 |
| Section: Abstract +1.5, Conclusion +1.5, Introduction +1, Results +1, Discussion +1, Limitations +0.5, Related work −1, Methods −1.5 | by section |
| First sentence of a paragraph, or a list item | +0.5 each |
The weights are a judgment, so Methods sentences with numbers are not promoted to claims. The map keeps the top-scoring candidates. It aims for about one claim per 40 words, within a target of 8 to 20 claims, and shows fewer when the text has fewer eligible sentences. Ties go to the earlier sentence. Each claim card shows its score and each point it received.
Evidence strength is a text rating from 0 to 3. It depends on what a sentence contains.
| Strength | Label | Rule in this tool |
|---|---|---|
| 0 | none | No number, citation, table pointer or method descriptor. |
| 1 | citation or assertion only | A citation such as [3], a figure or table pointer, or a method or dataset word such as "participants". |
| 2 | descriptive number | A number, without a test, interval or effect size. |
| 3 | number with test/interval/effect size | A number together with p, t(, F(, χ², r =, d =, OR =, RR =, CI, confidence interval, ±, or the words effect size. |
Linking uses BM25, a ranking formula that scores how well each evidence sentence matches the claim's words. The abstract is not used as evidence, because it usually restates the claims. A link needs at least two shared content words, and those words must cover at least a quarter of the claim's content words. Links are ordered by a BM25 score plus bonuses: 1 for the same paragraph, 0.5 for an adjacent paragraph, and 0.5 for a Results or Discussion sentence. BM25 depends on how many evidence sentences a text has, so the cover rule decides whether a link is made and the score only orders the links. Each claim shows at most three links, each with the shared words and the score.
A claim's strength is the highest strength among its links. A claim with no link is marked "unsupported in this text". That phrase describes this text only.
An overclaim flag appears when a claim uses strong or absolute wording and neither the claim sentence nor any linked evidence contains a number or a citation.
Statistics notes check four patterns. A p-value with no effect size or interval, and a test result with no sample size, are checked against the whole paragraph. A p-value below .05 phrased as proof, and a non-significant result described as no effect, are checked sentence by sentence. Each note links its source. Three of the four also name the ASA principle they apply.
The summary uses TextRank, a graph-based method that ranks sentences by how central they are to the rest of the text (Mihalcea and Tarau, TextRank, 2004). It picks up to 3, 5 or 8 sentences from the paper. It does not write new sentences.
Limits and what the map cannot tell you
Hedging is not the same as weakness. Hyland's book on hedging in research articles describes hedging as central to scientific argument. Its catalogue abstract says the study is based on journal articles and interviews with research scientists (Hyland, Hedging in Scientific Research Articles, 1998). UNC also says qualifiers are often necessary when evidence is open to doubt. So a hedged claim is not flagged as a problem. The map only shows the balance of hedged and absolute wording.
Rules miss things. A claim written without any cue word is not found. A sentence that a line break splits in an odd place can be read wrongly. A link can be made on shared vocabulary when the two sentences do not really concern the same thing. Open each linked sentence before you rely on a link.
The reading method behind the map comes from Keshav's guide to reading papers. He suggests up to three passes, and a first pass of about five to ten minutes. These times are his own experience, not measured values (Keshav, How to Read a Paper). The map is closest to his first pass: it finds the main claims and where their support is.
For a checklist of the same signs by hand, see How to spot weak evidence in a research paper. SciFact is a different task. It is a dataset of expert-written scientific claims, used to select research abstracts that support or refute each claim (Wadden et al., Fact or Fiction, EMNLP 2020). The map does not search other papers and does not check a claim against outside sources.
Reading a paper with the map
Start with the stat strip and the unsupported filter. A claim with no link is a question for the authors, not a verdict. Then open the paper pane and read the highlighted sentences in order.
The map weights sections differently. Abstract, Introduction, Results, Discussion and Conclusion sentences can be claims, and Methods and Related work sentences score lower. Any sentence with a number, a test, a citation, a figure or table pointer, or a method or dataset word can serve as evidence, except in the abstract, which usually restates the claims. This split is a design choice of this tool, not a rule any source states. The sources explain what the sections are for. UNC's scientific reports guide says Results gives "raw (i.e., uninterpreted) data" and Discussion "considers whether the data you obtained support the hypothesis" (UNC Writing Center, Scientific Reports). ICMJE asks authors to "avoid unqualified statements and conclusions not adequately supported by the data" (ICMJE). The overclaim flag checks for that kind of unqualified statement.
Use the list view for reading and the map for seeing the shape of the evidence. Both show the same claims and links. Copy the Markdown table to share the findings, or the anchors to find the same sentences again later.
If your paper is a PDF you will search again and again, PDF to Knowledge Pack turns PDFs into a searchable base with per-page citations. To see how a paper's text behaves when it is split for retrieval, Chunking Strategy Comparator compares seven strategies on your own text.
Privacy: what happens to your text
The tool's code reads and scores your text in this browser tab. It never sends the text to a server and never stores it. The tool keeps nothing in your browser, because there is no save option. The text stays in the page until you clear it or leave the page.
The PDF reader is a local copy of the PDF.js library, loaded from this site only after you choose a PDF. Exports are files made in your browser and downloaded to your computer.
The site-wide promo banner is a separate script on this page, not part of the tool. The tool's code does not pass your text to it. The banner keeps its own settings in your browser.
Questions
Is my data uploaded?
No. The tool's code reads and scores your text in this browser tab and never sends it to a server. The tool never stores your text, and it keeps nothing in your browser, because there is no save option. The PDF reader loads from this site only when you choose a PDF. The site-wide promo banner is a separate script, and the tool does not pass your text to it. Exports are files made in your browser.
How accurate is this?
It measures wording and structure, so accuracy means something narrower than it does for a fact checker. Claim detection and evidence linking use rules and word matching, and both make mistakes: a claim can be missed, or a link can be weak. Read each linked sentence before you rely on a flag. The tool cannot judge whether any claim is true.
Does an unsupported claim mean the paper is wrong?
No. Unsupported in this text means that no evidence sentence in the pasted text shares enough terms with the claim. The evidence may sit in a table, an appendix or another paper. A claim can be correct and still unsupported here, and a claim with a linked sentence can still be false.
Does p < .05 prove a finding?
No. The American Statistical Association says a p-value does not measure the probability that the studied hypothesis is true, and that conclusions should not rest only on whether a p-value passes a threshold (ASA press release, 2016). The map flags phrasing that treats p < .05 as proof. It does not check the test itself.
Does hedged wording make a paper weak?
Not by itself. UNC's Writing Center says writers need to clearly indicate whether claims are certain, likely, unlikely or just false, and that qualifiers are often necessary when evidence is open to doubt. Too many qualifiers can make a writer sound unsure. The map counts hedges so you can see the balance. It does not score the paper.
Why is a claim unlinked when a number sits right next to it?
A link needs at least two shared content words, and those words must cover at least a quarter of the claim's content words. A number with no shared words will not link, because the tool cannot tell which quantity belongs to which claim. Open the claim's anchor in the paper pane and read the sentences around it.
Can I use a PDF?
Yes, if the PDF has a text layer. The reader extracts text page by page, joins hyphenated words broken at a line end and drops bare page numbers. Scanned pages have no text layer and need OCR first. Each anchor carries its page number, so you can find the sentence in the original file. The limit is 80 pages.
Is the summary written by an AI?
No. The summary picks sentences that already appear in the paper, using TextRank, a graph-based ranking method. The same text always gives the same sentences. TextRank finds what is central in the text, which is not the same as what is correct.
Does this check claims against the research literature?
No. SciFact is a dataset for a different task: selecting research abstracts that support or refute a claim (Wadden et al., EMNLP 2020). This tool does not search other papers and does not check claims against outside sources. It reads only the text you paste.
Read more about this
Two plain-language guides that go deeper than this page can, with sources.