Guideline-Related Terms in IEEE VIS Papers, 1990–2024

This tab asks, for each exact occurrence, does this word name guidance produced by the current paper’s research contribution? Everything is computed over the 3,877-paper IEEE VIS corpus (1990–2024). Use the guidance views for this verified authored-guidance signal and Overall usage for the plain prevalence baseline.

What it means. How widely a term appears in the corpus at all, in any sense, before any judgement about meaning.

How it is calculated. A paper is counted once if it contains the term in any sense:

papers containing the term ÷ 3,877 corpus papers × 100

This is the same measure as “Overall use” in the Trends tab.

What it means. Appearances where the matched term itself names design or practice guidance that the current paper authors, derives, or explicitly owns as part of its findings, interviews, analysis, design work, or stated contribution.

How each appearance is classified. Every exact occurrence is evaluated separately. The classifier uses document location and local context, but a positive result requires both a guidance role and a direct grammatical link between that matched word and the current paper’s work:

1. Guidance-adjacent: the matched word directly names guidance produced by the current paper (for example, we derive design principles from our interviews).
2. Ambiguous / not verified: the phrase sounds like guidance or occurs in current-paper prose, but the matched word is not directly established as the paper’s authored guidance. Bare headings such as Design Implications remain here.
3. Likely non-guidance: prior or related work, citations and titles, general principles, idioms, technical senses, and system-generated outputs.

1. Guidance reach (default view): the absolute number of papers that use the term in a guidance-adjacent sense (prevalence combined with share).
2. Guidance share: of a term’s appearances, the percentage that are guidance-adjacent, on either basis:

   Per mention: guidance-adjacent occurrences ÷ all occurrences of the term × 100
Per paper:   papers with ≥1 guidance-adjacent use ÷ papers using the term × 100

3. Grid: the guidance-adjacent / ambiguous / non-guidance split for every term at once, as a colour grid.

For each term, let x be papers with at least one guidance-adjacent use, n be all papers containing the term, p̂ = x/n, and z = 1.96 for 95% confidence. No bootstrap or yearly averaging is used.

d = 1 + z²/n
center = (p̂ + z²/(2n)) / d
half-width = z × √(p̂(1−p̂)/n + z²/(4n²)) / d
interval = center ± half-width

Worked example. For principle, x = 31 and n = 1,328. The observed share is 2.3%; the Wilson calculation gives 1.6%–3.3% after rounding to one decimal place.

What it means. The bar is the descriptive classified share for this indexed corpus. The interval is a model-based indication of how stable that proportion is when these papers are treated as evidence of a broader publication pattern. Under repeated comparable samples, about 95% of intervals calculated this way would contain the underlying proportion.

Share and reach come apart: a term can have a relatively high verified share but occur in few papers, or be common overall while rarely naming guidance produced by the current paper.

1. Document location comes from TEI-like tags inside each raw foundText snippet. Tags were read before the snippet was cleaned for display, so citation and heading evidence did not get lost.
2. The phrase match itself still uses the same term families: principle(s), consideration(s), implication(s), rule(s) of thumb, and guardrail(s).
3. Only snippets with a recognized document-location tag are included in this normalized chart. The percentage is therefore "of recognized document-location evidence", not "of every possible extracted occurrence".

1. Body paragraph: <p>. This is authors' running prose and is the strongest location evidence that the current paper is using the phrase directly.
2. Reference list entry: <ref>. This is bibliography text. A phrase here usually belongs to a cited source, not the current paper's argument.
3. Title of cited work: <title> inside reference-like text. This catches cases where the phrase appears because another paper/book title contains it, such as Principles and Practice.
4. Section heading: <head>. Headings such as Design Implications or Design Considerations are useful structural signals, even when the heading is not a full sentence.
5. Figure/caption text: <figDesc>. This is near figures and captions rather than in main prose; it can be useful, but it is interpreted separately from body paragraphs.

1. Reference entries and cited titles are kept visible instead of silently removed because they explain why some terms have high raw counts but weaker guideline evidence.
2. For example, principle often appears in citation/title contexts, while consideration has a larger body-prose share.
3. Guardrail is especially sparse here: in this dataset its recognized document-location evidence appears in references or cited titles, not in body paragraphs.

1. Usage Types starts from the same matched snippets as Signal, then separates each term into more specific phrase-level categories.
2. The first pass checks document location: citation/title contexts are separated before deciding whether a phrase is design guidance, ambiguous, or likely noise.
3. The second pass checks the words immediately around the match: modifiers before the term, prepositions after it, and common fixed phrases such as in principle or under consideration.

1. Design or guideline principle: design principle(s), guiding principle(s), principles for..., or author prose where principles guide a design/evaluation choice.
2. Perception principle: named perceptual rules, especially Gestalt principle, proximity principle, common fate, or similar perception-language cues.
3. In principle: the fixed idiom in principle, meaning theoretically. This is classified as non-guidance.
4. Technical/mathematical: algorithmic or model-selection principles such as minimum description length principle.
5. Citation/publication title: reference/title contexts, especially books or papers whose titles contain Principles.

1. Design or practice consideration: design consideration(s), practical consideration(s), ethical consideration(s), accessibility consideration(s), or similar factors designers should account for.
2. Salient consideration: phrases such as key consideration, important consideration, main consideration, and central consideration.
3. Factor accounted for: phrasing like taken into consideration. This can be relevant, but it is ambiguous unless the snippet says what design action follows.
4. Object under study: under consideration, as in the algorithm under consideration or the dataset under consideration. This is non-guidance.
5. Technical consideration: practical implementation factors such as performance, latency, scalability, robustness, or memory constraints.

1. Design or practice implication: design implication(s), implications for design, implications for practice, or statements that connect findings to designer/practitioner action.
2. General consequence/relevance: has implications for... or broader implications without a concrete recommendation. These remain ambiguous.
3. Research/future implication: implications framed mainly as future work or future research.
4. Formal logic/fuzzy logic: technical uses such as logical implication or fuzzy implication, classified as non-guidance.

1. Design/visualization heuristic: a practical rule of thumb that advises a design choice, such as avoiding clutter or choosing a visual encoding.
2. Statistical/methodological rule: named formulas or conventions, including Silverman's rule of thumb, Freedman-Diaconis, Sturges, IQR, bandwidth rules, and bin-width rules.
3. Design/practice guardrail: a constraint, intervention, or boundary intended to steer visualization, dashboard, or analysis behavior.
4. AI/LLM guardrail: safety or collaboration constraints around model behavior or human-AI workflows.