Start with Keshav’s three passes
Keshav’s “How to Read a Paper” recommends reading in passes rather than front to back. In his words, “The key idea is that you should read the paper in up to three passes”. The first pass covers the title, abstract, introduction, section headings and conclusions. It ends with five questions about category, context, correctness, contributions and clarity.
The second pass looks at the figures, including whether the axes are labelled and whether the results show error bars. The third pass means re-creating the work and challenging each of its assumptions. Keshav also gives time budgets: “This pass should take about five to ten minutes” for the first pass, up to an hour for the second, and four or five hours for a beginner’s third pass, or about an hour for an experienced reader. These are his own experience-based estimates, not measured values.
The procedure below covers the first two passes. The third pass is work a text map cannot do for you.
Find the three to six main claims
A claim is a statement the author asks you to accept: a finding, a comparison or a conclusion. Start with the abstract, the contribution sentences in the introduction and the conclusion. Write each main claim in your own words, one per line, and stop at six. The tool returns its own candidates, so your six are a reading step, not its output. If you have more than six, you are probably recording supporting sentences as if they were claims.
Link each claim to the sentences that support it
For each claim, search the Results, figure captions and Discussion for sentences that bear on it, and keep up to three per claim. Count a sentence only if it reports something about the claim’s own subject. A methods sentence that contains a number describes a procedure; it is not a result.
The section a sentence sits in matters. UNC’s writing centre describes the Results row of the IMRaD format as one that provides raw (i.e., uninterpreted) data collected, while the Discussion considers whether the data you obtained support the hypothesis. Treat Results sentences as candidate evidence and Discussion sentences as interpretation to check.
Label each piece of evidence with one of four types, or None.
| Label | What it looks like | How much it carries |
|---|---|---|
| Number with a test or interval | A p-value, a t or F statistic, a confidence interval or an effect size | The strongest of the four labels, but a p-value alone says nothing about the size of an effect |
| Descriptive number | A mean, count or percentage with no test attached | Shows what was observed, not how much the result could vary |
| Figure or table pointer | “(Figure 2)” or “Table 3 shows…”, with the numbers only in the figure | Real evidence, but you must open the figure and check its axes and error bars |
| Citation or assertion only | “consistent with earlier work [7]” or “this is well known” | Points to someone else’s evidence or to the author’s word |
| None | No sentence in the text bears on the claim | Unsupported in this text. That does not make the claim false |
Mark wording strength and note what is missing
Next, mark how strongly each claim is worded. Hedged words such as may, might, could and suggests tell the reader the author is unsure. Absolute words such as always, never and all tell the reader the author is certain. UNC’s qualifiers page says writers need to show whether they think a claim is certain, likely, unlikely, or just false, and notes that qualifiers are often necessary when the evidence is open to doubt.
Hedging is not a weakness in itself. The same page warns that excessive use of qualifiers can make you sound unsure of your facts. So read wording against evidence rather than alone. A strong verb backed by a test is a different case from the same verb backed only by a citation.
Then note what is missing. Check each claim for these gaps:
- Scope: the claim covers more cases than the study sampled.
- Size: a p-value with no effect size or interval beside it.
- Sample: no sample size near the result.
- Limits: the Limitations section does not mention the gap.
A worked example from an invented abstract
The excerpt below is invented: a fictional study of commuters in the fictional city of Harlow Bay, not a real paper. Each line carries an anchor label.
§Abstract ¶1 s1 We studied 412 commuters in the fictional city of Harlow Bay.
§Abstract ¶1 s2 Protected bike lanes always reduce commute times.
§Abstract ¶1 s3 These results suggest that lanes may encourage more people to cycle.
§Results ¶1 s1 Riders on the lane had commutes 4.1 minutes shorter on average (t(410) = 2.3, p = .02).
§Results ¶1 s2 Median car commutes did not change (Figure 2).
§Results ¶2 s1 Lane use was not significantly associated with weekly exercise (p = .21).
§Discussion ¶1 s1 The effect is consistent with earlier city studies [7].
§Discussion ¶2 s1 Our results show that the lane shortens every commute.
§Limitations ¶1 s1 Route choice and cycling frequency were self-reported.
The table maps five claims by hand. It is a reading of the invented text, not tool output.
| Claim (anchor) | Wording | Linked evidence and type | What is missing |
|---|---|---|---|
| “Protected bike lanes always reduce commute times.” (§Abstract ¶1 s2) | Absolute: “always” | §Results ¶1 s1, number with a test | Any city besides Harlow Bay. Nothing in the excerpt matches “always”. |
| “Median car commutes did not change.” (§Results ¶1 s2) | Plain; a negative result | Same sentence, figure pointer | The values sit in Figure 2. Check its axes and error bars. |
| “Lane use was not significantly associated with weekly exercise.” (§Results ¶2 s1) | Plain; a negative result | Same sentence, number with a test | No effect size or interval in the sentence. |
| “These results suggest that lanes may encourage more people to cycle.” (§Abstract ¶1 s3) | Hedged: “suggest”, “may” | Hand reading: §Discussion ¶1 s1, citation only. The tool does not select this sentence as a claim. | Cycling frequency was self-reported (§Limitations ¶1 s1). The excerpt gives no figure for it. |
| “Our results show that the lane shortens every commute.” (§Discussion ¶2 s1) | Strong “show that”, absolute “every” | None that covers every commute | Per-rider data. The results report an average of 4.1 minutes. |
The tool’s own run on this text differs from rows 4 and 5. It returns two claims, the abstract’s “always” sentence and the Discussion’s “show that … every” sentence. It links both to §Results ¶1 s1 on the shared words “lane” and “commutes”, and it does not select the hedged sentence §Abstract ¶1 s3 as a claim. That link is lexical: the Results sentence reports an average for riders, not every commute. Row 1 would not trip the tool’s overclaim flag, because its linked evidence contains a test, yet the claim reaches well beyond one city. Row 3 counts as a number with a test, but a p-value does not measure size. The American Statistical Association puts it directly: “A p-value, or statistical significance, does not measure the size of an effect or the importance of a result.”
Run the same steps locally in the Claim–Evidence Map
The Claim–Evidence Map applies a similar procedure in your browser. It never sends or stores your paper. Paste a paper’s text, or load its PDF, up to 80 pages. It picks claim sentences from contribution wording such as “we show”, weighting abstract and conclusion sentences over methods. For each claim it links up to three sentences that share its key terms, and it leaves a claim unlinked when none clears its threshold.
Every sentence gets a stable label, such as §Results ¶3 s2, plus a page number for PDFs such as p. 4. Cite those labels in your notes: the same text gives the same labels on every run. Each claim also gets a wording label (hedged, moderate, strong or absolute) with the matched words shown. Evidence is scored on four levels: none, citation or assertion only, descriptive number, and number with a test, interval or effect size. The tool scores a figure pointer at the same level as a citation, so check figures by hand.
A claim with no linked evidence is shown as unsupported in this text. That wording is deliberate. The tool sees only the text you give it, so unsupported means no linked sentence was found in the text, not that the claim is false. If you want to search a PDF’s pages with citations, the PDF to Knowledge Pack builds a searchable index with page references.
What a claim map cannot tell you
The Claim–Evidence Map reads wording and structure. It cannot tell you whether a claim is true, whether the methods were sound or whether the sample represents anyone. Its links are word matches: a sentence can be linked because it shares vocabulary with the claim, which is not the same as supporting it. The ASA’s statement concludes with a wider caution: “No single index should substitute for scientific reasoning”.
Common mistakes:
- Reading a strength label or overclaim flag as a quality score for the whole paper.
- Reading “unsupported in this text” as “wrong”.
- Trusting a claim because it appears in the abstract, where summaries can overstate the Results.
- Counting every sentence with a number as evidence, including methods descriptions.
For the next step, read the companion guide on how to spot weak evidence in a research paper.
Common questions
How long does a claim–evidence map take?
The tool publishes no timing figure. Keshav’s time budgets are his own experience with reading papers, not measurements of any tool.
Does a claim–evidence map show whether a paper is true?
No. It shows which claims have linked sentences in the text and how strongly each claim is worded. Judging truth needs the methods, the wider literature and your own third pass.
What is the difference between a claim and evidence?
A claim is a statement the author asks you to accept. Evidence is what bears on it: a number with a test, a descriptive number, a figure pointer or a citation.
Is a hedged claim always weaker?
No. UNC’s writing centre says qualifiers are often necessary when evidence is open to doubt. Compare each hedge with the evidence behind it. A strong verb over a citation alone is the pattern to question.
What does “unsupported in this text” mean?
It means no sentence in the text you pasted bears on the claim strongly enough to be linked. It does not mean the claim is false, and the evidence may sit in a section you did not paste.
Can I run this on a PDF?
Yes, up to 80 pages. The tool reads the PDF’s text layer, so a scanned paper with no text layer gives it nothing to read.
Sources and further reading
- Keshav, “How to Read a Paper” — S. Keshav, David R. Cheriton School of Computer Science, University of Waterloo, undated
- Wasserstein and Lazar, “The ASA’s Statement on p-Values: Context, Process, and Purpose” — The American Statistician, vol. 70, no. 2, 2016 (archived copy)
- American Statistical Association, statement on statistical significance and p-values — ASA release, 7 March 2016
- UNC Writing Center, “Qualifiers” — University of North Carolina at Chapel Hill, undated
- UNC Writing Center, “Scientific Reports” — University of North Carolina at Chapel Hill, undated