Skip to main content

9. Figures and tables that work

Many reviewers read your figures and tables before a single sentence of prose, and form their opinion there. Figure 1, usually your architecture or main idea, and Table 1, usually your headline result, persuade more than any paragraph. Methods tells you to spend a full day on Figure 1; Results tells you the main table is the headline. This chapter shows how to make figures and tables that are correct, legible, accessible, and ACL-compliant, and which tools to use. Apply one test throughout: a reader who sees only your figures, tables and captions should be able to reconstruct the paper.

What ACL requires​

Get these wrong and you look careless to the people deciding your fate. From the ACL formatting guidelines:

  • Use vector graphics. "Graphics and photos should, if possible, use vector graphic formats (PDF, EPS), which allow the graphics to scale arbitrarily." A blurry PNG screenshot of a plot is the most common amateur tell. Export PDF.
  • Keep the font large. "For any text or numbers in tables and figures, whenever possible, please use the font size of the document text." Reviewers are irritated by six-point axis labels nobody can read.
  • Captions are 10-point and self-contained. The ACL guide places them below the figure or table, but follow the year's style files. Make every caption a short paragraph that explains the figure without the body text (Results).
  • It must survive print and zoom. "Your paper must look good both when printed (A4 size) and when viewed onscreen as PDF." Check both.
  • Grayscale-readable. "To accommodate people who are color-blind (as well as those printing with black-and-white printers), grayscale readability is strongly encouraged… tables and figures [must] not rely solely on color to convey critical distinctions."

Color, done right​

Roughly eight percent of men have some color-vision deficiency, and plenty of reviewers still print in black and white. Never make color your only signal.

  • Use a colorblind-safe palette. Good defaults: viridis, the Okabe–Ito palette, ColorBrewer schemes, or seaborn's "colorblind". Do not separate categories by red versus green alone.
  • Encode redundantly. Pair color with line style, marker shape or a direct label, so the figure still works in grayscale. Aim for WCAG AA contrast (4.5:1 for normal text).
  • Test it. Run the figure through a simulator (Color Oracle, or the Coblis web tool) and print it in grayscale. If two lines merge, redesign before a reviewer points it out.

Plots that work​

Rougier et al.'s "Ten Simple Rules for Better Figures" (PLOS Computational Biology, 2014, open access) is the classic short reference. Claus Wilke's free Fundamentals of Data Visualization goes deeper. The essentials:

  • One figure, one message. Decide the single point the plot must make, then strip everything that does not serve it. A figure that tries to show five things shows none of them.
  • Maximize data, minimize ink. Tufte's principle: no 3-D bars, no gradient fills, no heavy gridlines, no chartjunk. Let the data dominate the frame.
  • Label honestly. Give axes units. Use a legend or, better, direct labels on the lines. Add error bars or shaded variance where you have multiple seeds.
  • Do not mislead with scale. Start bar charts at zero, never truncate an axis to inflate a gap, and use the same axis range across plots you want compared.
  • Tools. matplotlib + seaborn (Python) or ggplot2 (R). Save with a vector backend (plt.savefig("fig.pdf", bbox_inches="tight")) and a colorblind palette. Generate the plot from your results files, so it updates when the numbers do.

Architecture and method diagrams​

Figure 1 is often the most reproduced part of a paper, so give it real time. A reader should be able to explain your method from the diagram alone.

  • Show flow. Lay it out left to right or top down so the data path is obvious. Use arrows for data flow, and the same shape for the same kind of thing.
  • Label the tensors. Name inputs, outputs and key dimensions, so a reader sees where the shapes change.
  • Cut hard. A diagram with forty boxes communicates nothing. Show the idea, and move the exhaustive detail to the text or an appendix figure.
  • Tools. TikZ/PGF is the publication standard: native LaTeX, vector, beautiful, with a steep learning curve. draw.io / diagrams.net and Excalidraw are fast, free and export PDF. Inkscape and Graphviz suit specific needs. Keynote or PowerPoint are fine if you export vector PDF instead of a screenshot. Whatever you draw in, export vector.

Tables that work​

A good results table is dense, scannable and honest. In LaTeX, the booktabs package is the fix that pays most.

  • No vertical rules. Use booktabs' \toprule, \midrule, \bottomrule, with no vertical lines and no double rules. The eye has to fight through them.
  • Bold the best per column. If the difference is not significant, bold both, so you do not imply a winner the statistics do not support (Experimental Setup).
  • Show variance and align numbers. Report 78.4 ± 0.6, keep significant figures consistent down a column, put units in the header, and align on the decimal point (the siunitx S column does this for you).
  • Right-align numbers, left-align text, group related rows, and separate with whitespace instead of lines.
  • Keep the giant table out of the body. A 40-column monster belongs in the appendix. The main text gets the rows that carry your claims.
  • Tools. booktabs + siunitx for the styling. Generate the table body from code (pandas df.to_latex() or df.style, or a small script), so the numbers in the paper never drift from the numbers you computed.

Accessibility is now expected​

*ACL venues ask for accessible papers, and reviewers increasingly notice. Beyond the colorblind-safe palette and redundant encoding above, the ACL accessibility guidance asks for alt text on every visual: figures, tables, charts and diagrams alike. Write alt text that states the salient finding ("Accuracy rises with model size on all three languages, fastest for Hausa"). A flat description ("a line chart") is not enough, and neither is a copy of the caption. Treat alt text as part of the work.

Tools at a glance​

JobReach for
Statistical plotsmatplotlib + seaborn, ggplot2, plotly
Colorblind-safe colorviridis, Okabe–Ito, ColorBrewer, seaborn colorblind
Architecture diagramsTikZ/PGF, draw.io, Excalidraw, Inkscape, Graphviz
LaTeX tablesbooktabs, siunitx, tabularray, pandas to_latex
Checking color / grayscaleColor Oracle, Coblis, print it in B&W

A short workflow​

  • Draft the key figure and table early. Deciding what Figure 1 and Table 1 must show forces you to pin down your claim.
  • Generate from data, never hand-draw numbers, so everything updates when results change.
  • Export vector, embed fonts, and check the figure at 100% and zoomed in.
  • Run the grayscale and colorblind check, then write the alt text.
  • Read each caption alone. If it does not stand without the body text, it is not finished.

Common mistakes​

  • Raster screenshots. A blurry PNG of a plot or a table. Export vector PDF.
  • Tiny fonts. Axis labels and table numbers smaller than the body text.
  • Color as the only signal, or a red-green palette. Encode redundantly; use a safe palette.
  • Chartjunk and misleading axes. 3-D bars, truncated y-axes, mismatched scales across compared plots.
  • Vertical rules and double lines in tables. Use booktabs.
  • Label-only captions, and figures or tables the prose never refers to.
  • No alt text. Reviewers now expect it in an accessible submission.

Further reading​