Colour in Translation: Data, Models, and Benchmarking for Cross-Linguistic Colour Naming

Paper Title

Colour in Translation: Data, Models, and Benchmarking for Cross-Linguistic Colour Naming

Publication Info

• Topic area: Cross-linguistic colour naming, computational modelling, and machine translation.
• Keywords: Colour naming, cross-linguistic variation, computational models, machine translation, Spin Colour Forest, perceptual grounding, multilingual datasets, benchmarking, human-computer interaction, cultural diversity.

Background and Problem

• Problem / challenge: Effective cross-linguistic colour communication is hindered by the lack of multilingual datasets, computational models, and benchmarks for colour name translation. Prior models often rely on universal Basic Colour Terms (BCTs), which fail to capture the richer, language-specific partitioning of colour space.
• Significance: Accurate colour naming and translation are essential for human-computer interaction (HCI) tasks such as interface design, data visualisation, and cross-cultural communication, where misinterpretation can lead to exclusion or inefficiency.
• Motivation and related work: Previous studies, including the World Colour Survey and computational models, focused on constrained naming systems (e.g., BCTs) and single-language datasets. These approaches fail to account for linguistic diversity and perceptual differences across cultures. This paper builds on these foundations by addressing the gaps in multilingual datasets, modelling, and benchmarking.

Solution

• Proposed approach: Spin Colour Forest (SCF), a novel ensemble regression method that applies partial orthogonal transformations to perceptual colour space to model cross-linguistic colour naming distributions.
• Novelty:
  1. Creation of unconstrained multilingual datasets for five languages (American English, British English, French, Greek, Himba) with 70,052 responses to 600 colour stimuli.
  2. Introduction of SCF, which captures complex colour category boundaries using geometric transformations in perceptual colour space.
  3. Development of the first systematic colour translation benchmark to evaluate lexical and perceptual accuracy in AI translation systems.
• Procedure and key techniques:
  • Data collection via online experiments (English, French, Greek) and field studies (Himba) using 600 colour stimuli in CAM16-UCS space.
  • Preprocessing responses to normalize spelling, remove noise, and identify unique colour terms.
  • Training SCF models with partial orthogonal transformations to predict colour naming distributions.
  • Identifying indispensable colour vocabularies for each language using systematic grid sampling.
  • Establishing translation benchmarks using Jensen-Shannon divergence (JSD) and human validation.

Results

• Concrete findings:
  • Indispensable colour vocabularies identified: 47 terms for American English, 32 for British English, 27 for French, 32 for Greek, and 7 for Himba.
  • SCF improved Colour Naming Fidelity (CNF) scores by 2.7–4.4% across languages compared to baseline models.
  • Lexical translation accuracy (chrF++) was highest for English variants (up to 78.99) and poorest for Himba (28.90), with perceptual accuracy (ΔE) showing large errors across all languages (average 17–28 units).
• Advantage over baselines:
  • SCF outperformed Random Forest and Extra Trees models in CNF scores and valid name coverage.
  • Transformation-enhanced methods captured complex linguistic colour naming boundaries better than unrotated models.
• Experiments / evaluation:
  • Evaluation of SCF using 3-fold cross-validation, systematic test grids, and statistical comparisons.
  • Translation benchmark tested Claude Sonnet 4 LLM for lexical and perceptual accuracy across five languages.
• Limitations and future work:
  • Limited language scope (five languages); future expansion to underrepresented regions and languages planned.
  • Evaluation focused on literal colour naming, not figurative or contextual uses.
  • Further validation across diverse LLM architectures and broader datasets required.

Summary

This study advances cross-linguistic colour naming research by introducing multilingual datasets, the Spin Colour Forest model, and a systematic translation benchmark. SCF effectively models colour naming distributions across five languages, identifying indispensable vocabularies that surpass traditional Basic Colour Terms. Translation benchmarks reveal critical limitations in large language models, highlighting a lexical-perceptual disconnect that undermines cross-cultural colour communication. These findings provide actionable insights for HCI practitioners, emphasizing the need for culturally aware AI-powered tools and interfaces. The datasets, models, and benchmarks released enable future research and applications in inclusive design and multilingual systems.