To translate a scientific figure correctly, preserve the measurement, unit, comparison, uncertainty and scope of its claim. Scientific graph translation, axis-label localization, figure legend translation and error-bar terminology are not cosmetic editing tasks. A wrong denominator, an unexplained logarithmic axis or a switch from standard deviation to standard error can change what a reader believes the experiment showed. The aim is a faithful scientific representation in natural target-language prose, not a graph that merely looks polished.
This guide explains how to translate charts in journal articles, research posters, preprints, theses and classroom materials. It covers axes, legends, panel labels, error bars, sample sizes, significance markers, colours, uncertainty intervals, normalized quantities and accessibility descriptions. The method is designed for translators, students, researchers, publishers and editors working across languages, using fictional examples rather than pretending to reproduce unprovided datasets.
Begin with the figure as a system. The graphic encodes values and relationships; the caption defines symbols and procedures; the surrounding Results and Methods supply limits. Translate those layers together. If an axis is ‘growth rate (cm/day)’ and the target quietly becomes ‘growth (cm)’, a careful sentence has concealed a changed variable. If error bars are labelled ‘mean ± SEM’ but the translation says ‘range’, the reader is no longer seeing the same evidence.
The visible problem: an elegant translated caption can misreport the study
Imagine two line graphs displayed side by side. Both have identical curves, but one y-axis is a linear concentration scale and the other is logarithmic. Translate the axis label and ignore the scale, and a reader may infer that equal visual distances represent equal arithmetic changes. In another figure, two bars have error caps, but the caption identifies them as 95% confidence intervals. Rendering those as standard deviations does more than change a term: it changes the statistical quantity being presented.
Scientific figures are densely compressed arguments. Their language includes units, labels, abbreviations, sample definitions, numerical signs, asterisks and sometimes visually encoded exclusions. Translators who have only the caption should ask for the figure and linked manuscript text. Translators who have only an exported SVG should ask for the source legend, because visual observation cannot establish which statistic an error bar represents.
Fifteen first principles of translating quantitative graphics
1. Preserve the measured variable
Concentration, rate, count, proportion, probability, change and normalized value are distinct. A broad word like ‘level’ may hide what was calculated. The objective is to preserve a variable and its evidence, not to make all terms sound familiar to nonspecialists. Scientific precision and readable English or another language are compatible when the translator understands the underlying distinction.
For instance: Growth rate in mm/day cannot be translated as final length in mm. The correct render may require a parenthetical definition, but it should not invent a result or erase necessary qualifications. When the source itself is ambiguous, request information rather than choosing a convenient statistic.
A useful audit question is: Identify the numerator, denominator and what each plotted observation denotes. The answer belongs in a bilingual figure map that the editor can verify independently.
2. Carry units with quantities
Numerical labels are meaningless without their unit, scale and dimensional context. Units belong to the variable, not to decorative typesetting. The objective is to preserve a variable and its evidence, not to make all terms sound familiar to nonspecialists. Scientific precision and readable English or another language are compatible when the translator understands the underlying distinction.
For instance: A rate of 20 µg/mL is not the same as 20 mg/mL. The correct render may require a parenthetical definition, but it should not invent a result or erase necessary qualifications. When the source itself is ambiguous, request information rather than choosing a convenient statistic.
A useful audit question is: Check unit symbols, powers and denominators character by character. The answer belongs in a bilingual figure map that the editor can verify independently.
3. Keep linear and log scales distinct
Equal positions on a logarithmic scale represent multiplicative, not additive, steps. Translation must retain explicit scale notes. The objective is to preserve a variable and its evidence, not to make all terms sound familiar to nonspecialists. Scientific precision and readable English or another language are compatible when the translator understands the underlying distinction.
For instance: A log10 concentration axis is not an ordinary concentration scale. The correct render may require a parenthetical definition, but it should not invent a result or erase necessary qualifications. When the source itself is ambiguous, request information rather than choosing a convenient statistic.
A useful audit question is: Read tick marks and the Methods description before assigning a scale. The answer belongs in a bilingual figure map that the editor can verify independently.
4. Respect transformations
Log-transformed, standardized, scaled, centered, normalized and indexed values are not direct raw measurements. The objective is to preserve a variable and its evidence, not to make all terms sound familiar to nonspecialists. Scientific precision and readable English or another language are compatible when the translator understands the underlying distinction.
For instance: A plot of z scores must not become a plot of raw test marks. The correct render may require a parenthetical definition, but it should not invent a result or erase necessary qualifications. When the source itself is ambiguous, request information rather than choosing a convenient statistic.
A useful audit question is: Record the mathematical transformation and reference baseline. The answer belongs in a bilingual figure map that the editor can verify independently.
5. Define the centre statistic
Mean, median and geometric mean differ. Some figures show model estimates rather than raw group averages. The objective is to preserve a variable and its evidence, not to make all terms sound familiar to nonspecialists. Scientific precision and readable English or another language are compatible when the translator understands the underlying distinction.
For instance: Median with interquartile range cannot be rewritten as mean with standard deviation. The correct render may require a parenthetical definition, but it should not invent a result or erase necessary qualifications. When the source itself is ambiguous, request information rather than choosing a convenient statistic.
A useful audit question is: Find the centre statistic in the caption and Methods. The answer belongs in a bilingual figure map that the editor can verify independently.
6. Do not interchange SD SEM and CI
Standard deviation describes variation; standard error concerns uncertainty of an estimate under assumptions; a confidence interval is an inferential interval with a stated construction. The objective is to preserve a variable and its evidence, not to make all terms sound familiar to nonspecialists. Scientific precision and readable English or another language are compatible when the translator understands the underlying distinction.
For instance: ‘Mean ± SD’ and ‘mean ± SEM’ should never be merged into ‘average with error’. The correct render may require a parenthetical definition, but it should not invent a result or erase necessary qualifications. When the source itself is ambiguous, request information rather than choosing a convenient statistic.
A useful audit question is: Match every error bar to its exact named measure. The answer belongs in a bilingual figure map that the editor can verify independently.
7. Name the observational unit
Biological replicates, technical repeats, subjects, cultures and experimental runs may give very different meanings to n. The objective is to preserve a variable and its evidence, not to make all terms sound familiar to nonspecialists. Scientific precision and readable English or another language are compatible when the translator understands the underlying distinction.
For instance: Three wells from one biological sample may not constitute three independent participants. The correct render may require a parenthetical definition, but it should not invent a result or erase necessary qualifications. When the source itself is ambiguous, request information rather than choosing a convenient statistic.
A useful audit question is: Preserve what n counts and whether observations are independent. The answer belongs in a bilingual figure map that the editor can verify independently.
8. Keep comparisons and groups aligned
Group A, control, treated, baseline, female/male categories and age bands require faithful identity. The objective is to preserve a variable and its evidence, not to make all terms sound familiar to nonspecialists. Scientific precision and readable English or another language are compatible when the translator understands the underlying distinction.
For instance: The control group cannot be renamed reference treatment if that suggests an intervention. The correct render may require a parenthetical definition, but it should not invent a result or erase necessary qualifications. When the source itself is ambiguous, request information rather than choosing a convenient statistic.
A useful audit question is: Trace legend labels to their actual curves and colours. The answer belongs in a bilingual figure map that the editor can verify independently.
9. P values are not effect sizes
A reported p value is associated with a statistical procedure and assumptions. It does not directly state the size or importance of an effect. The objective is to preserve a variable and its evidence, not to make all terms sound familiar to nonspecialists. Scientific precision and readable English or another language are compatible when the translator understands the underlying distinction.
For instance: ‘p = 0.03’ must not become ‘3% improvement’. The correct render may require a parenthetical definition, but it should not invent a result or erase necessary qualifications. When the source itself is ambiguous, request information rather than choosing a convenient statistic.
A useful audit question is: Check whether p is exact, thresholded, adjusted or linked to a specific test. The answer belongs in a bilingual figure map that the editor can verify independently.
10. Retain inequality signs and thresholds
Less than, greater than, at most and at least are different relationships. A small glyph error can reverse a result. The objective is to preserve a variable and its evidence, not to make all terms sound familiar to nonspecialists. Scientific precision and readable English or another language are compatible when the translator understands the underlying distinction.
For instance: p < 0.001 is not p = 0.001, and ≤ differs from <. The correct render may require a parenthetical definition, but it should not invent a result or erase necessary qualifications. When the source itself is ambiguous, request information rather than choosing a convenient statistic.
A useful audit question is: Compare operators against the source rather than rewriting from memory. The answer belongs in a bilingual figure map that the editor can verify independently.
11. Panel labels are navigational references
Figure 2a, 2b and 2c establish which subfigure supports which sentence. Their order must stay linked. The objective is to preserve a variable and its evidence, not to make all terms sound familiar to nonspecialists. Scientific precision and readable English or another language are compatible when the translator understands the underlying distinction.
For instance: If panel B shows dose and panel C shows time, the captions cannot be swapped. The correct render may require a parenthetical definition, but it should not invent a result or erase necessary qualifications. When the source itself is ambiguous, request information rather than choosing a convenient statistic.
A useful audit question is: Check every cross-reference in the article and supplementary files. The answer belongs in a bilingual figure map that the editor can verify independently.
12. Captions are not a place to invent inference
A figure legend can describe what is plotted, methods and statistical symbols; translation must not add claims absent from the source. The objective is to preserve a variable and its evidence, not to make all terms sound familiar to nonspecialists. Scientific precision and readable English or another language are compatible when the translator understands the underlying distinction.
For instance: ‘Difference observed’ should not become ‘treatment proved effective’. The correct render may require a parenthetical definition, but it should not invent a result or erase necessary qualifications. When the source itself is ambiguous, request information rather than choosing a convenient statistic.
A useful audit question is: Separate description, interpretation and speculation. The answer belongs in a bilingual figure map that the editor can verify independently.
13. Colour is one encoding among several
Series may be distinguished by colour, shape, pattern or line style. Translating a colour name alone can fail for colour-blind or nonvisual readers. The objective is to preserve a variable and its evidence, not to make all terms sound familiar to nonspecialists. Scientific precision and readable English or another language are compatible when the translator understands the underlying distinction.
For instance: A red dashed curve and blue solid curve require accurate identity even in monochrome print. The correct render may require a parenthetical definition, but it should not invent a result or erase necessary qualifications. When the source itself is ambiguous, request information rather than choosing a convenient statistic.
A useful audit question is: Check line styles and legend descriptions in grayscale. The answer belongs in a bilingual figure map that the editor can verify independently.
14. Accessibility should retain relationships
A complex chart often needs a short alternative and a longer accessible description or data table. Translating visual appearance alone loses the results. The objective is to preserve a variable and its evidence, not to make all terms sound familiar to nonspecialists. Scientific precision and readable English or another language are compatible when the translator understands the underlying distinction.
For instance: ‘Four coloured bars’ tells a screen reader user almost nothing about comparative quantities. The correct render may require a parenthetical definition, but it should not invent a result or erase necessary qualifications. When the source itself is ambiguous, request information rather than choosing a convenient statistic.
A useful audit question is: Test whether the reader receives the axes, groups, trend and limitations. The answer belongs in a bilingual figure map that the editor can verify independently.
15. Scientific names and symbols need discipline
Latin binomials, gene symbols, chemical formulas, Greek characters and isotope labels can carry exact identities. Translators need domain-aware formatting. The objective is to preserve a variable and its evidence, not to make all terms sound familiar to nonspecialists. Scientific precision and readable English or another language are compatible when the translator understands the underlying distinction.
For instance: The gene ABC1 is not necessarily equivalent to a common-language abbreviation sharing those letters. The correct render may require a parenthetical definition, but it should not invent a result or erase necessary qualifications. When the source itself is ambiguous, request information rather than choosing a convenient statistic.
A useful audit question is: Preserve identifiers and check specialized nomenclature rules. The answer belongs in a bilingual figure map that the editor can verify independently.
The twelve-stage scientific figure translation workflow
Stage 1: Collect every layer of the figure
Obtain image, editable source if possible, caption, Results paragraph, Methods, data dictionary and supplementary notes.
Example: A legend that says ‘a.u.’ may be explained in Methods, not in the exported PNG. The figure translator needs both a linguistic and scientific explanation for each decision, and must preserve the source evidence even where a stylistic editor prefers a shorter phrase.
Deliverable: A single figure dossier with version number. Reopen the actual graphic to verify it; a spreadsheet-only review cannot establish whether curves, panels and values remained aligned.
Stage 2: Identify the figure’s actual question
Write the comparison the figure is designed to answer and the narrowest conclusion the data permit.
Example: One graph may describe an association, not an intervention’s causal effect. The figure translator needs both a linguistic and scientific explanation for each decision, and must preserve the source evidence even where a stylistic editor prefers a shorter phrase.
Deliverable: A neutral one-sentence evidence summary. Reopen the actual graphic to verify it; a spreadsheet-only review cannot establish whether curves, panels and values remained aligned.
Stage 3: Map axes and scales
Record variable names, units, minimum and maximum, transformation, linear/log scale, tick conventions and direction.
Example: An axis labelled ’10^3 cells’ is a scaling convention, not a new unit of cell identity. The figure translator needs both a linguistic and scientific explanation for each decision, and must preserve the source evidence even where a stylistic editor prefers a shorter phrase.
Deliverable: A source-target axis dictionary. Reopen the actual graphic to verify it; a spreadsheet-only review cannot establish whether curves, panels and values remained aligned.
Stage 4: Map all series and panels
Record panel letters, groups, colours, shapes, line types, time points and legend cross-references.
Example: A dotted blue line may stand for projection; a solid blue line may stand for observations. The figure translator needs both a linguistic and scientific explanation for each decision, and must preserve the source evidence even where a stylistic editor prefers a shorter phrase.
Deliverable: A verified panel and series mapping. Reopen the actual graphic to verify it; a spreadsheet-only review cannot establish whether curves, panels and values remained aligned.
Stage 5: Extract statistical definitions
Identify centre measure, spread or interval type, exact n, statistical test, sidedness and multiple-testing adjustments if stated.
Example: An asterisk legend may denote a threshold only for selected comparisons. The figure translator needs both a linguistic and scientific explanation for each decision, and must preserve the source evidence even where a stylistic editor prefers a shorter phrase.
Deliverable: A statistics checklist linked to the manuscript. Reopen the actual graphic to verify it; a spreadsheet-only review cannot establish whether curves, panels and values remained aligned.
Stage 6: Lock numbers and symbols
Protect numerals, decimals, p-values, signs, Greek letters, units, exponents, scientific names and mathematical operators.
Example: Changing a decimal separator or minus sign can alter an experimental result. The figure translator needs both a linguistic and scientific explanation for each decision, and must preserve the source evidence even where a stylistic editor prefers a shorter phrase.
Deliverable: A locked-token list for each figure. Reopen the actual graphic to verify it; a spreadsheet-only review cannot establish whether curves, panels and values remained aligned.
Stage 7: Translate natural labels
Render concise variable labels and explanatory phrases in natural target language while keeping defined abbreviations consistent.
Example: ‘Relative abundance’ may need a brief definition when the readership is broad. The figure translator needs both a linguistic and scientific explanation for each decision, and must preserve the source evidence even where a stylistic editor prefers a shorter phrase.
Deliverable: A termbase with approved target phrases. Reopen the actual graphic to verify it; a spreadsheet-only review cannot establish whether curves, panels and values remained aligned.
Stage 8: Reconstruct the caption
Write the title, panels, displayed quantity, replicates, error-bar definition, significance notes and exceptions in the correct sequence.
Example: A merged multi-panel figure can conceal differing sample sizes. The figure translator needs both a linguistic and scientific explanation for each decision, and must preserve the source evidence even where a stylistic editor prefers a shorter phrase.
Deliverable: A self-contained caption checked against all panels. Reopen the actual graphic to verify it; a spreadsheet-only review cannot establish whether curves, panels and values remained aligned.
Stage 9: Test the statistical meaning
Ask whether a reader would understand the same uncertainty and claim, not simply the same familiar vocabulary.
Example: Overlap of two bars does not by itself settle a significance claim. The figure translator needs both a linguistic and scientific explanation for each decision, and must preserve the source evidence even where a stylistic editor prefers a shorter phrase.
Deliverable: A second person’s statistical review or a documented query. Reopen the actual graphic to verify it; a spreadsheet-only review cannot establish whether curves, panels and values remained aligned.
Stage 10: Review layout and legibility
Check expansion, line breaks, small text, subscript/superscript, font support, RTL scripts and printed scale.
Example: An exported image can clip a translated unit or lose the superscript in m². The figure translator needs both a linguistic and scientific explanation for each decision, and must preserve the source evidence even where a stylistic editor prefers a shorter phrase.
Deliverable: A layout proof at final intended dimensions. Reopen the actual graphic to verify it; a spreadsheet-only review cannot establish whether curves, panels and values remained aligned.
Stage 11: Provide an accessible equivalent
Translate short alternatives, long descriptions and data tables as needed, retaining relationships and exclusions.
Example: A complex scatterplot with a fitted line may require textual trend and uncertainty notes. The figure translator needs both a linguistic and scientific explanation for each decision, and must preserve the source evidence even where a stylistic editor prefers a shorter phrase.
Deliverable: An accessibility review separate from visual proofing. Reopen the actual graphic to verify it; a spreadsheet-only review cannot establish whether curves, panels and values remained aligned.
Stage 12: Sign off source-to-target fidelity
Verify the final chart and caption after production rather than relying on translation-memory snippets.
Example: Editors may re-export labels or swap versions after the language review. The figure translator needs both a linguistic and scientific explanation for each decision, and must preserve the source evidence even where a stylistic editor prefers a shorter phrase.
Deliverable: A recorded final check of published artifact. Reopen the actual graphic to verify it; a spreadsheet-only review cannot establish whether curves, panels and values remained aligned.
Twenty-eight worked examples: labels, uncertainty and conclusions
These are constructed learning examples, not a reanalysis of any published study. Each identifies an English gloss of a hypothetical source graphic, a tempting but misleading translation, a working target phrasing and an evidence check. A real translation must also preserve the original target-language units and editorial style.
Worked example 1: Rate versus amount
Hypothetical source: “Y-axis ‘CO₂ uptake (µmol m⁻² s⁻¹)’” A misleading translation might say “CO₂ quantity (µmol)” The surface wording may look close, but it changes how a scientist, student or reader would interpret the graph. The error could arise in an axis, figure key, caption, table, abstract or accessible description.
A better working rendering is “CO₂ uptake rate (µmol m⁻² s⁻¹)” Rationale: The area and time denominators define a rate, not an amount. Preserve the measurement and its limitations even where the result becomes less dramatic or requires one more explanatory phrase. Accuracy matters more than a shorter, more marketable caption.
Verification: Are both superscript denominators intact? Compare the final language version with the displayed scale, statistical notation and surrounding Results section. If the underlying source is internally inconsistent, record an author query instead of repairing the data silently.
Worked example 2: Count versus concentration
Hypothetical source: “Plot labels ‘cells per mL’” A misleading translation might say “number of cells” The surface wording may look close, but it changes how a scientist, student or reader would interpret the graph. The error could arise in an axis, figure key, caption, table, abstract or accessible description.
A better working rendering is “cells per millilitre” Rationale: Removing the volume denominator changes what is being compared. Preserve the measurement and its limitations even where the result becomes less dramatic or requires one more explanatory phrase. Accuracy matters more than a shorter, more marketable caption.
Verification: Does the source measurement use a fixed volume? Compare the final language version with the displayed scale, statistical notation and surrounding Results section. If the underlying source is internally inconsistent, record an author query instead of repairing the data silently.
Worked example 3: Percent and percentage points
Hypothetical source: “Prevalence rises from 10% to 15%” A misleading translation might say “increased by 5%” The surface wording may look close, but it changes how a scientist, student or reader would interpret the graph. The error could arise in an axis, figure key, caption, table, abstract or accessible description.
A better working rendering is “increased by 5 percentage points” Rationale: An absolute percentage-point difference is not the same as relative percent change. Preserve the measurement and its limitations even where the result becomes less dramatic or requires one more explanatory phrase. Accuracy matters more than a shorter, more marketable caption.
Verification: Does text refer to absolute or relative change? Compare the final language version with the displayed scale, statistical notation and surrounding Results section. If the underlying source is internally inconsistent, record an author query instead of repairing the data silently.
Worked example 4: Logarithmic ticks
Hypothetical source: “Y-axis log10 copies per gram” A misleading translation might say “copies per gram on linear scale” The surface wording may look close, but it changes how a scientist, student or reader would interpret the graph. The error could arise in an axis, figure key, caption, table, abstract or accessible description.
A better working rendering is “log10 copies per gram” Rationale: The transformation changes interpretation of equal visual spaces. Preserve the measurement and its limitations even where the result becomes less dramatic or requires one more explanatory phrase. Accuracy matters more than a shorter, more marketable caption.
Verification: Are tick labels powers of ten? Compare the final language version with the displayed scale, statistical notation and surrounding Results section. If the underlying source is internally inconsistent, record an author query instead of repairing the data silently.
Worked example 5: Negative values
Hypothetical source: “Difference from baseline: -2.4” A misleading translation might say “2.4 higher than baseline” The surface wording may look close, but it changes how a scientist, student or reader would interpret the graph. The error could arise in an axis, figure key, caption, table, abstract or accessible description.
A better working rendering is “2.4 below baseline” Rationale: The negative sign gives the direction of change. Preserve the measurement and its limitations even where the result becomes less dramatic or requires one more explanatory phrase. Accuracy matters more than a shorter, more marketable caption.
Verification: Is the sign clearly visible? Compare the final language version with the displayed scale, statistical notation and surrounding Results section. If the underlying source is internally inconsistent, record an author query instead of repairing the data silently.
Worked example 6: SD bars
Hypothetical source: “Mean ± SD, n=12” A misleading translation might say “mean with standard error” The surface wording may look close, but it changes how a scientist, student or reader would interpret the graph. The error could arise in an axis, figure key, caption, table, abstract or accessible description.
A better working rendering is “mean ± standard deviation; n = 12” Rationale: The target must preserve the spread measure chosen by the author. Preserve the measurement and its limitations even where the result becomes less dramatic or requires one more explanatory phrase. Accuracy matters more than a shorter, more marketable caption.
Verification: Does the caption define n correctly? Compare the final language version with the displayed scale, statistical notation and surrounding Results section. If the underlying source is internally inconsistent, record an author query instead of repairing the data silently.
Worked example 7: SEM bars
Hypothetical source: “Mean ± SEM” A misleading translation might say “mean ± standard deviation” The surface wording may look close, but it changes how a scientist, student or reader would interpret the graph. The error could arise in an axis, figure key, caption, table, abstract or accessible description.
A better working rendering is “mean ± standard error of the mean” Rationale: SEM is not interchangeable with SD and should not be inferred from the bar’s width. Preserve the measurement and its limitations even where the result becomes less dramatic or requires one more explanatory phrase. Accuracy matters more than a shorter, more marketable caption.
Verification: Can Methods confirm SEM? Compare the final language version with the displayed scale, statistical notation and surrounding Results section. If the underlying source is internally inconsistent, record an author query instead of repairing the data silently.
Worked example 8: Confidence intervals
Hypothetical source: “95% CI around estimated risk ratio” A misleading translation might say “95% of observations fall here” The surface wording may look close, but it changes how a scientist, student or reader would interpret the graph. The error could arise in an axis, figure key, caption, table, abstract or accessible description.
A better working rendering is “95% confidence interval for the estimated risk ratio” Rationale: The interval concerns an estimated parameter under a statistical procedure, not a count of observations. Preserve the measurement and its limitations even where the result becomes less dramatic or requires one more explanatory phrase. Accuracy matters more than a shorter, more marketable caption.
Verification: Is the parameter and CI construction documented? Compare the final language version with the displayed scale, statistical notation and surrounding Results section. If the underlying source is internally inconsistent, record an author query instead of repairing the data silently.
Worked example 9: Sample size
Hypothetical source: “n = 8 independent experiments” A misleading translation might say “eight cells were measured” The surface wording may look close, but it changes how a scientist, student or reader would interpret the graph. The error could arise in an axis, figure key, caption, table, abstract or accessible description.
A better working rendering is “eight independent experiments” Rationale: The observational unit matters to the validity of comparisons. Preserve the measurement and its limitations even where the result becomes less dramatic or requires one more explanatory phrase. Accuracy matters more than a shorter, more marketable caption.
Verification: Does n refer to experiments, participants or fields? Compare the final language version with the displayed scale, statistical notation and surrounding Results section. If the underlying source is internally inconsistent, record an author query instead of repairing the data silently.
Worked example 10: Technical replicates
Hypothetical source: “Triplicate wells per sample” A misleading translation might say “three independent samples” The surface wording may look close, but it changes how a scientist, student or reader would interpret the graph. The error could arise in an axis, figure key, caption, table, abstract or accessible description.
A better working rendering is “three technical measurements per sample” Rationale: Replication of measurement is not automatically biological independence. Preserve the measurement and its limitations even where the result becomes less dramatic or requires one more explanatory phrase. Accuracy matters more than a shorter, more marketable caption.
Verification: Are biological n and technical n separately shown? Compare the final language version with the displayed scale, statistical notation and surrounding Results section. If the underlying source is internally inconsistent, record an author query instead of repairing the data silently.
Worked example 11: Median and IQR
Hypothetical source: “Median [IQR]” A misleading translation might say “mean and standard deviation” The surface wording may look close, but it changes how a scientist, student or reader would interpret the graph. The error could arise in an axis, figure key, caption, table, abstract or accessible description.
A better working rendering is “median and interquartile range” Rationale: Centre and spread were both changed in the unsafe version. Preserve the measurement and its limitations even where the result becomes less dramatic or requires one more explanatory phrase. Accuracy matters more than a shorter, more marketable caption.
Verification: Are box boundaries explained? Compare the final language version with the displayed scale, statistical notation and surrounding Results section. If the underlying source is internally inconsistent, record an author query instead of repairing the data silently.
Worked example 12: Geometric mean
Hypothetical source: “Geometric mean fluorescence intensity” A misleading translation might say “ordinary average intensity” The surface wording may look close, but it changes how a scientist, student or reader would interpret the graph. The error could arise in an axis, figure key, caption, table, abstract or accessible description.
A better working rendering is “geometric mean fluorescence intensity” Rationale: A geometric mean can be appropriate on different scales and is not an arithmetic mean. Preserve the measurement and its limitations even where the result becomes less dramatic or requires one more explanatory phrase. Accuracy matters more than a shorter, more marketable caption.
Verification: Are transformed values involved? Compare the final language version with the displayed scale, statistical notation and surrounding Results section. If the underlying source is internally inconsistent, record an author query instead of repairing the data silently.
Worked example 13: Adjusted p value
Hypothetical source: “Adjusted p = 0.042” A misleading translation might say “uncorrected p 0.042” The surface wording may look close, but it changes how a scientist, student or reader would interpret the graph. The error could arise in an axis, figure key, caption, table, abstract or accessible description.
A better working rendering is “multiple-comparison-adjusted p = 0.042” Rationale: Omitting adjustment can misrepresent the statistical method. Preserve the measurement and its limitations even where the result becomes less dramatic or requires one more explanatory phrase. Accuracy matters more than a shorter, more marketable caption.
Verification: What correction does Methods specify? Compare the final language version with the displayed scale, statistical notation and surrounding Results section. If the underlying source is internally inconsistent, record an author query instead of repairing the data silently.
Worked example 14: Inequality preservation
Hypothetical source: “p < 0.001” A misleading translation might say “p = 0.001” The surface wording may look close, but it changes how a scientist, student or reader would interpret the graph. The error could arise in an axis, figure key, caption, table, abstract or accessible description.
A better working rendering is “p less than 0.001” Rationale: The original reports a bound, not an exact value. Preserve the measurement and its limitations even where the result becomes less dramatic or requires one more explanatory phrase. Accuracy matters more than a shorter, more marketable caption.
Verification: Was the sign copied accurately? Compare the final language version with the displayed scale, statistical notation and surrounding Results section. If the underlying source is internally inconsistent, record an author query instead of repairing the data silently.
Worked example 15: Non-significant result
Hypothetical source: “No evidence of a difference at threshold” A misleading translation might say “the two groups are equal” The surface wording may look close, but it changes how a scientist, student or reader would interpret the graph. The error could arise in an axis, figure key, caption, table, abstract or accessible description.
A better working rendering is “no statistically detected difference under the stated test” Rationale: Failure to detect is not proof of equivalence. Preserve the measurement and its limitations even where the result becomes less dramatic or requires one more explanatory phrase. Accuracy matters more than a shorter, more marketable caption.
Verification: What confidence bounds were reported? Compare the final language version with the displayed scale, statistical notation and surrounding Results section. If the underlying source is internally inconsistent, record an author query instead of repairing the data silently.
Worked example 16: Panel reference
Hypothetical source: “Fig. 3b shows velocity; Fig. 3c shows acceleration” A misleading translation might say “Fig. 3c velocity; 3b acceleration” The surface wording may look close, but it changes how a scientist, student or reader would interpret the graph. The error could arise in an axis, figure key, caption, table, abstract or accessible description.
A better working rendering is “Panel b: velocity; panel c: acceleration” Rationale: Reversing panel letters changes where the reader looks for evidence. Preserve the measurement and its limitations even where the result becomes less dramatic or requires one more explanatory phrase. Accuracy matters more than a shorter, more marketable caption.
Verification: Do all in-text references agree? Compare the final language version with the displayed scale, statistical notation and surrounding Results section. If the underlying source is internally inconsistent, record an author query instead of repairing the data silently.
Worked example 17: Projection versus observation
Hypothetical source: “Solid = observed; dashed = modelled” A misleading translation might say “two measured series” The surface wording may look close, but it changes how a scientist, student or reader would interpret the graph. The error could arise in an axis, figure key, caption, table, abstract or accessible description.
A better working rendering is “solid observed; dashed modelled” Rationale: An estimate or forecast is not an observation. Preserve the measurement and its limitations even where the result becomes less dramatic or requires one more explanatory phrase. Accuracy matters more than a shorter, more marketable caption.
Verification: Is the legend’s line style preserved? Compare the final language version with the displayed scale, statistical notation and surrounding Results section. If the underlying source is internally inconsistent, record an author query instead of repairing the data silently.
Worked example 18: Normalization baseline
Hypothetical source: “Relative expression, control = 1” A misleading translation might say “absolute expression one unit” The surface wording may look close, but it changes how a scientist, student or reader would interpret the graph. The error could arise in an axis, figure key, caption, table, abstract or accessible description.
A better working rendering is “expression normalized to control value of 1” Rationale: The normalization explains why the control appears at one. Preserve the measurement and its limitations even where the result becomes less dramatic or requires one more explanatory phrase. Accuracy matters more than a shorter, more marketable caption.
Verification: Was the denominator specified? Compare the final language version with the displayed scale, statistical notation and surrounding Results section. If the underlying source is internally inconsistent, record an author query instead of repairing the data silently.
Worked example 19: Per 100,000 rate
Hypothetical source: “Cases per 100,000 people” A misleading translation might say “raw case counts” The surface wording may look close, but it changes how a scientist, student or reader would interpret the graph. The error could arise in an axis, figure key, caption, table, abstract or accessible description.
A better working rendering is “cases per 100,000 population” Rationale: Population scaling is part of the variable. Preserve the measurement and its limitations even where the result becomes less dramatic or requires one more explanatory phrase. Accuracy matters more than a shorter, more marketable caption.
Verification: Is the time period also stated? Compare the final language version with the displayed scale, statistical notation and surrounding Results section. If the underlying source is internally inconsistent, record an author query instead of repairing the data silently.
Worked example 20: Standardized z score
Hypothetical source: “Standardized score (z)” A misleading translation might say “raw score” The surface wording may look close, but it changes how a scientist, student or reader would interpret the graph. The error could arise in an axis, figure key, caption, table, abstract or accessible description.
A better working rendering is “z score relative to stated reference distribution” Rationale: A standardized variable cannot be read in original units. Preserve the measurement and its limitations even where the result becomes less dramatic or requires one more explanatory phrase. Accuracy matters more than a shorter, more marketable caption.
Verification: Which reference mean and SD were used? Compare the final language version with the displayed scale, statistical notation and surrounding Results section. If the underlying source is internally inconsistent, record an author query instead of repairing the data silently.
Worked example 21: Scale bar
Hypothetical source: “Micrograph scale bar 20 μm” A misleading translation might say “20× magnification” The surface wording may look close, but it changes how a scientist, student or reader would interpret the graph. The error could arise in an axis, figure key, caption, table, abstract or accessible description.
A better working rendering is “scale bar: 20 micrometres” Rationale: Scale bars give length, not simply optical magnification. Preserve the measurement and its limitations even where the result becomes less dramatic or requires one more explanatory phrase. Accuracy matters more than a shorter, more marketable caption.
Verification: Is the bar inside the image consistent with caption? Compare the final language version with the displayed scale, statistical notation and surrounding Results section. If the underlying source is internally inconsistent, record an author query instead of repairing the data silently.
Worked example 22: Asterisk legend
Hypothetical source: “* p < 0.05; ** p < 0.01” A misleading translation might say “asterisks mean strong result” The surface wording may look close, but it changes how a scientist, student or reader would interpret the graph. The error could arise in an axis, figure key, caption, table, abstract or accessible description.
A better working rendering is “one and two asterisks denote the stated p thresholds” Rationale: Symbols are coding conventions, not independent evidence of practical importance. Preserve the measurement and its limitations even where the result becomes less dramatic or requires one more explanatory phrase. Accuracy matters more than a shorter, more marketable caption.
Verification: Which exact comparison does each symbol mark? Compare the final language version with the displayed scale, statistical notation and surrounding Results section. If the underlying source is internally inconsistent, record an author query instead of repairing the data silently.
Worked example 23: Missing data
Hypothetical source: “Excluded 4 missing observations” A misleading translation might say “all data included” The surface wording may look close, but it changes how a scientist, student or reader would interpret the graph. The error could arise in an axis, figure key, caption, table, abstract or accessible description.
A better working rendering is “four observations excluded due to missingness” Rationale: Exclusion affects the denominator and interpretation. Preserve the measurement and its limitations even where the result becomes less dramatic or requires one more explanatory phrase. Accuracy matters more than a shorter, more marketable caption.
Verification: Are reasons and n updated? Compare the final language version with the displayed scale, statistical notation and surrounding Results section. If the underlying source is internally inconsistent, record an author query instead of repairing the data silently.
Worked example 24: Error bands
Hypothetical source: “Shaded region denotes 95% CI” A misleading translation might say “region shows observed data range” The surface wording may look close, but it changes how a scientist, student or reader would interpret the graph. The error could arise in an axis, figure key, caption, table, abstract or accessible description.
A better working rendering is “shaded region represents a 95% confidence interval” Rationale: An inferential band is not the same as the raw data envelope. Preserve the measurement and its limitations even where the result becomes less dramatic or requires one more explanatory phrase. Accuracy matters more than a shorter, more marketable caption.
Verification: Does a model or measured mean lie at its centre? Compare the final language version with the displayed scale, statistical notation and surrounding Results section. If the underlying source is internally inconsistent, record an author query instead of repairing the data silently.
Worked example 25: Two y axes
Hypothetical source: “Left axis temperature; right axis humidity” A misleading translation might say “both curves show temperature” The surface wording may look close, but it changes how a scientist, student or reader would interpret the graph. The error could arise in an axis, figure key, caption, table, abstract or accessible description.
A better working rendering is “left: temperature; right: relative humidity” Rationale: A double-axis figure needs explicit series-to-axis mapping. Preserve the measurement and its limitations even where the result becomes less dramatic or requires one more explanatory phrase. Accuracy matters more than a shorter, more marketable caption.
Verification: Can a reader recover units for both? Compare the final language version with the displayed scale, statistical notation and surrounding Results section. If the underlying source is internally inconsistent, record an author query instead of repairing the data silently.
Worked example 26: Odds ratio
Hypothetical source: “Odds ratio 1.8 (95% CI 1.1–3.0)” A misleading translation might say “risk increases 80%” The surface wording may look close, but it changes how a scientist, student or reader would interpret the graph. The error could arise in an axis, figure key, caption, table, abstract or accessible description.
A better working rendering is “odds ratio 1.8 with 95% confidence interval 1.1–3.0” Rationale: Odds and risk are not identical measures; avoid translating one as the other. Preserve the measurement and its limitations even where the result becomes less dramatic or requires one more explanatory phrase. Accuracy matters more than a shorter, more marketable caption.
Verification: Was the original measure OR, RR or HR? Compare the final language version with the displayed scale, statistical notation and surrounding Results section. If the underlying source is internally inconsistent, record an author query instead of repairing the data silently.
Worked example 27: Correlation versus causation
Hypothetical source: “Observed correlation r = 0.62” A misleading translation might say “treatment caused increase” The surface wording may look close, but it changes how a scientist, student or reader would interpret the graph. The error could arise in an axis, figure key, caption, table, abstract or accessible description.
A better working rendering is “observed positive correlation, r = 0.62” Rationale: A statistical association is not automatically causal evidence. Preserve the measurement and its limitations even where the result becomes less dramatic or requires one more explanatory phrase. Accuracy matters more than a shorter, more marketable caption.
Verification: Does study design support causal inference? Compare the final language version with the displayed scale, statistical notation and surrounding Results section. If the underlying source is internally inconsistent, record an author query instead of repairing the data silently.
Worked example 28: Illustrative model fit
Hypothetical source: “Line shows model prediction with CI” A misleading translation might say “the line is the actual measurements” The surface wording may look close, but it changes how a scientist, student or reader would interpret the graph. The error could arise in an axis, figure key, caption, table, abstract or accessible description.
A better working rendering is “modelled prediction line and its stated confidence interval” Rationale: An estimated curve and individual observations belong to different evidence layers. Preserve the measurement and its limitations even where the result becomes less dramatic or requires one more explanatory phrase. Accuracy matters more than a shorter, more marketable caption.
Verification: Are points and prediction lines distinguished? Compare the final language version with the displayed scale, statistical notation and surrounding Results section. If the underlying source is internally inconsistent, record an author query instead of repairing the data silently.
What the word ‘uncertainty’ does and does not mean
Statistical uncertainty is not interchangeable with ordinary doubt. The standard deviation of observed values, the standard error of an estimated mean, and a confidence interval are different objects. An error bar could represent any of these, or an entirely different quantity. The translator must not infer its identity by looking at the figure’s visual width. The caption or Methods should define it, and absent definitions deserve a query.
A p value also does not tell a reader the probability that the scientific hypothesis is true. Its interpretation depends on the statistical model and test. Nor does a nonsignificant result establish that two treatments are identical. Keeping distinctions like these prevents a translation from turning measured evidence into a stronger claim than the study supports.
Twelve failure modes that change what the figure claims
1. Dropping scale words
A linear y-axis is silently substituted for a log scale. This is not merely a typographical error when it changes a result or experimental boundary. Include the transformation and validate the tick labels. Document the repair in a figure checklist and revisit the associated sentence in the main text, since mistakes often propagate between graphs and prose.
2. Changing units
Micrograms become milligrams while numeric values stay unchanged. This is not merely a typographical error when it changes a result or experimental boundary. Check units dimensionally and preserve symbols. Document the repair in a figure checklist and revisit the associated sentence in the main text, since mistakes often propagate between graphs and prose.
3. Flattening a rate
A per-day measurement becomes a total amount. This is not merely a typographical error when it changes a result or experimental boundary. Keep the denominator and study period. Document the repair in a figure checklist and revisit the associated sentence in the main text, since mistakes often propagate between graphs and prose.
4. Using generic ‘average’ for every centre
A median result becomes an arithmetic mean. This is not merely a typographical error when it changes a result or experimental boundary. Use the stated statistic and define it where needed. Document the repair in a figure checklist and revisit the associated sentence in the main text, since mistakes often propagate between graphs and prose.
5. Calling every error bar a range
A standard error or confidence interval becomes a raw data spread. This is not merely a typographical error when it changes a result or experimental boundary. Recover the exact definition from the legend or Methods. Document the repair in a figure checklist and revisit the associated sentence in the main text, since mistakes often propagate between graphs and prose.
6. Inventing a sample size
The editor assumes n denotes participants where it denotes images. This is not merely a typographical error when it changes a result or experimental boundary. Identify the real observational unit. Document the repair in a figure checklist and revisit the associated sentence in the main text, since mistakes often propagate between graphs and prose.
7. Swapping legend colours
Two series are assigned to the wrong groups. This is not merely a typographical error when it changes a result or experimental boundary. Check line shapes, series IDs and colours against the source. Document the repair in a figure checklist and revisit the associated sentence in the main text, since mistakes often propagate between graphs and prose.
8. Promoting p values to effect sizes
p = 0.02 becomes a 2% treatment gain. This is not merely a typographical error when it changes a result or experimental boundary. Keep hypothesis-test results separate from measured differences. Document the repair in a figure checklist and revisit the associated sentence in the main text, since mistakes often propagate between graphs and prose.
9. Omitting negative signs
An observed decline is reported as an increase. This is not merely a typographical error when it changes a result or experimental boundary. Lock minus signs and examine values visually. Document the repair in a figure checklist and revisit the associated sentence in the main text, since mistakes often propagate between graphs and prose.
10. Hiding exclusions
A figure loses its missing-data note. This is not merely a typographical error when it changes a result or experimental boundary. Preserve who or what is omitted and why. Document the repair in a figure checklist and revisit the associated sentence in the main text, since mistakes often propagate between graphs and prose.
11. Treating association as proof
The target caption claims causality not established by the study. This is not merely a typographical error when it changes a result or experimental boundary. Keep the design’s evidential limits visible. Document the repair in a figure checklist and revisit the associated sentence in the main text, since mistakes often propagate between graphs and prose.
12. Ignoring production changes
Late image export swaps panel letters after translation review. This is not merely a typographical error when it changes a result or experimental boundary. Perform a final asset-level bilingual comparison. Document the repair in a figure checklist and revisit the associated sentence in the main text, since mistakes often propagate between graphs and prose.
Guided practice and transfer to classrooms, papers and public reports
Practice task 1: Rate or amount?
Translate ‘evaporation rate (mm/day)’ and ‘cumulative evaporation (mm)’ as separate axis labels. Answer: one is time-normalized and the other is an accumulated amount. Keep both units and their distinct physical meaning.
For a rigorous exercise, ask one learner to translate the labels, another to reconstruct the data question from the target alone, and a third to verify the scientific meaning against the source. Record the disagreement precisely: symbol, unit, group, statistic, time period or claim strength. This method creates an observable check rather than a vague judgement that the translation sounds right.
Practice task 2: Which error bar?
A caption contains ‘mean ± SEM, n = 24’ but the graph template labels the key ‘SD’. Draft a translator query. Answer: do not pick the supposedly more common option; ask which statistic was computed and correct the figure and caption together.
For a rigorous exercise, ask one learner to translate the labels, another to reconstruct the data question from the target alone, and a third to verify the scientific meaning against the source. Record the disagreement precisely: symbol, unit, group, statistic, time period or claim strength. This method creates an observable check rather than a vague judgement that the translation sounds right.
Practice task 3: Log plot literacy
A figure uses a log10 concentration axis with ticks 1, 10, 100 and 1000. Explain the scale in another language without calling it a linear increase. Answer: equal graphical steps here indicate tenfold changes.
For a rigorous exercise, ask one learner to translate the labels, another to reconstruct the data question from the target alone, and a third to verify the scientific meaning against the source. Record the disagreement precisely: symbol, unit, group, statistic, time period or claim strength. This method creates an observable check rather than a vague judgement that the translation sounds right.
Practice task 4: Effect and significance
A paper reports a small group difference with p = 0.04. Draft two sentences, one descriptive and one cautious interpretive statement. Answer: keep the reported effect size distinct from the p value and avoid saying the treatment is certainly useful.
For a rigorous exercise, ask one learner to translate the labels, another to reconstruct the data question from the target alone, and a third to verify the scientific meaning against the source. Record the disagreement precisely: symbol, unit, group, statistic, time period or claim strength. This method creates an observable check rather than a vague judgement that the translation sounds right.
Practice task 5: Panel audit
A three-panel infographic has labels A, B and C, but its translated caption lists B, A, C. Draft an audit sheet. Answer: map each caption phrase to its image panel and the exact in-text citation, then correct all mismatches.
For a rigorous exercise, ask one learner to translate the labels, another to reconstruct the data question from the target alone, and a third to verify the scientific meaning against the source. Record the disagreement precisely: symbol, unit, group, statistic, time period or claim strength. This method creates an observable check rather than a vague judgement that the translation sounds right.
Practice task 6: Accessible figure summary
A chart compares measured and projected temperatures with shaded uncertainty. Draft short alt and a long-description outline. Answer: identify measured versus modelled, axes with units, main trend, shaded band definition and limitations, avoiding made-up data.
For a rigorous exercise, ask one learner to translate the labels, another to reconstruct the data question from the target alone, and a third to verify the scientific meaning against the source. Record the disagreement precisely: symbol, unit, group, statistic, time period or claim strength. This method creates an observable check rather than a vague judgement that the translation sounds right.
Frequently asked questions
How do I translate scientific graphs accurately?
Start by mapping axes, units, transformations, group labels, sample size and statistical definitions. Translate the caption with the graphic and Methods in view, then test every numeric and scientific label. If the source does not tell you enough to answer accurately, issue a specific query rather than substituting a familiar label. A good figure translation keeps the same scientific question answerable.
Can I translate axis titles without seeing the graph?
You can draft terminology, but not certify the figure. The scale, series and visual context may reveal a critical distinction or inconsistency. If the source does not tell you enough to answer accurately, issue a specific query rather than substituting a familiar label. A good figure translation keeps the same scientific question answerable.
Is standard deviation the same as standard error?
No. SD describes spread of observations, while SEM describes an aspect of uncertainty of an estimated mean under assumptions. Preserve the original statistic. If the source does not tell you enough to answer accurately, issue a specific query rather than substituting a familiar label. A good figure translation keeps the same scientific question answerable.
What does a 95% confidence interval mean?
It is an interval produced by a statistical procedure with a long-run coverage interpretation under its assumptions, not a guarantee that 95% of observations lie inside. If the source does not tell you enough to answer accurately, issue a specific query rather than substituting a familiar label. A good figure translation keeps the same scientific question answerable.
What does n mean in a figure?
It means the stated number of units counted, but which units matter: people, independent runs, images or technical repeats? Find the definition before translating. If the source does not tell you enough to answer accurately, issue a specific query rather than substituting a familiar label. A good figure translation keeps the same scientific question answerable.
Should I replace unfamiliar p-value symbols with words?
You may add explanatory wording for readers, but keep inequality signs, values, thresholds and adjustments intact. If the source does not tell you enough to answer accurately, issue a specific query rather than substituting a familiar label. A good figure translation keeps the same scientific question answerable.
Can a translator correct a suspicious figure?
Flag inconsistencies and propose questions, but do not silently amend data or statistical interpretations. The author or qualified analyst must approve scientific corrections. If the source does not tell you enough to answer accurately, issue a specific query rather than substituting a familiar label. A good figure translation keeps the same scientific question answerable.
How should I handle logarithmic axes?
Preserve the transformation and read scale ticks accurately. Do not write a linear-scale description for an axis with multiplicative spacing. If the source does not tell you enough to answer accurately, issue a specific query rather than substituting a familiar label. A good figure translation keeps the same scientific question answerable.
How do I translate captions for multi-panel figures?
Map panel letters and references first, then translate each panel explanation with its variables and statistical notes. Verify final layout after production. If the source does not tell you enough to answer accurately, issue a specific query rather than substituting a familiar label. A good figure translation keeps the same scientific question answerable.
Does a p value measure practical importance?
No. Reported statistical evidence and the magnitude or real-world importance of an effect are distinct. Avoid inserting an efficacy claim absent from the study. If the source does not tell you enough to answer accurately, issue a specific query rather than substituting a familiar label. A good figure translation keeps the same scientific question answerable.
Can accessibility alt text replace the caption?
Complex scientific graphics may need complementary short text and longer data description. The caption and accessible equivalent can support each other but have distinct purposes. If the source does not tell you enough to answer accurately, issue a specific query rather than substituting a familiar label. A good figure translation keeps the same scientific question answerable.
What is the fastest final audit?
Check every axis name, unit, sign, log transformation, series, panel letter, n value, centre statistic, uncertainty measure, p-value operator and conclusion against the original image and manuscript. If the source does not tell you enough to answer accurately, issue a specific query rather than substituting a familiar label. A good figure translation keeps the same scientific question answerable.
A final research-figure review checklist
Verify the figure version, all axes and labels, units and their powers, values and signs, normalization denominators, log or linear scale, colour and line-style mapping, panel letters, legend symbols, centre and spread measures, number and definition of replicates, p-value notation and significance legend. Check the figure’s main takeaway against the full surrounding Results paragraph, and inspect the final exported graphic rather than trusting the editable file.
Where accessibility is required, supply descriptive text and structured numerical access so the reader can understand the comparison without relying on colour. Do not make up missing values in a description. Track the named dataset, study population, measurement procedure and any exclusion stated by the authors.
Technical references and evidence
The Nature formatting guide directs authors to define error bars and statistics in figure legends. Krzywinski and Altman, “Error bars,” Nature Methods (2013) explains why error-bar types and overlap are easily misread. PLOS ONE submission guidelines address units, captions and statistical reporting. For accessible complex graphics, consult the W3C image alternatives decision tree. Follow the standards appropriate to the intended journal and discipline; these references are not a substitute for reviewing the actual manuscript.
Next routes inside the eduKateSG translation library
Revisit Translate Meaning, Not Just Words to distinguish the evidence from its wording. Use Alt Text and Screen Reader Labels for the separate job of accessible graphics descriptions. The series index routes specialist fields; Vocabulary Learning Hub develops fine distinctions; and How English Works: Translation and Meaning explains what precision must survive across languages.
The most useful translator question is not “Does this caption sound scientific?” It is “Could a researcher reconstruct the same measurement, comparison and degree of uncertainty from the target language?” When the answer is yes, language has transferred the evidence rather than merely the appearance of expertise.