Word–Color Association

Code accompanying the manuscript Word–Color Association from Large-Scale Online Data: Methodology and Applications by Sungpil Wang and Juyong Park.

The framework retrieves word–color associations from publicly accessible Google Image Search results: it scrapes the top-results screenshot for a query word, resizes it to 10,000 pixels, converts each pixel to the CIELch color space, and represents the word as a 10-dimensional additive ratio vector over named base colors (red, orange, yellow, green, blue, purple, pink, black, grey, white).

---

Installation

You can set up the required environment using either requirements.txt (pip) or environment.yml (conda).

Option 1 — pip

brew install python@3.9
/opt/homebrew/opt/python@3.9/bin/python3.9 -m venv venv
source venv/bin/activate
python -m ensurepip --upgrade
pip install --upgrade pip setuptools wheel
pip install -r requirements.txt

Option 2 — conda

conda env create -f environment.yaml
conda activate word-color-association

Chrome and chromedriver are required for word2image.py (Google Image Search screenshot via Selenium).

---

Quick start

Extract the color vector for a single word:

python main.py --term "moss" --option ratio

Visualize the color information of a target word on a local webpage:

cd website
python app.py

---

Core pipeline

• word2image.py — Google Image Search screenshot via Selenium (incognito, regional/language presets, light-mode forced via CDP)
• calculation.py — pixel-to-CIELch conversion, base-color binning, per-word 10-d vector computation
• color_utils.py — shared color-space utilities (rgb_to_lch, color_vector_from_pixels, resize_to_pixel_count, base-color ranges, chroma threshold)
• visualization.py — per-word color distribution plots
• main.py — command-line entry point

Notebooks

• Text2Color.ipynb — extract color vectors per word in long sentences and render as text + color-bar visualizations
• LogoAndScreenshotOfBrands.ipynb — top-100 publicly traded companies' logo vs. web-image color comparison
• ThresholdOfChroma.ipynb — chroma distribution of 15 achromatic terms and the chroma threshold derivation
• WhiteGridRatio.ipynb — white-divider trimming ratio derived from a 100-screenshot sample (Kolmogorov–Smirnov normality check)
• ComparisonPoets.ipynb — Lowell vs. Johnson corpus-level color comparison

---

External validation

Each script produces CSV / PDF / PNG / summary-MD outputs under csv/revision_results/.

• nrc_validation.py — NRC Word–Colour Association Lexicon (Mohammad 2013): per-color one-vs-rest AUC + confusion matrix
• rathore_scrape.py, rathore_compare.py — graded slider ratings (Rathore et al. 2019, IEEE TVCG): per-concept cosine on the UW-58 / BCP-37 palettes
• compsyn_compare.py — Representational Similarity Analysis (RSA) vs. comp-syn 8-bin JzAzBz histograms (Guilbeault et al. 2020)
• setlur_scrape.py, setlur_full_pipeline.py, setlur_analyze.py — Setlur & Stone (2016) full-pipeline reproduction: WordNet color-term lookup with Bing-Image-Search dominant-color extraction (k-means + W3C-nearest)

Analyses

• permutation_null.py — Lowell vs. Johnson Δ permutation null test (B = 10,000 size-matched partitions)
• weighting_sensitivity.py, alpha_robustness_report.py — α-sweep robustness of the Lowell–Johnson contrast over indirect-word weights
• polysemy_scrape.py, polysemy_demo.py — bare-token vs. full-registered-name color recovery for polysemous queries
• scroll_depth_scrape.py, page_compare_scrape.py, page_compare_analyze.py — page-1 vs. page-10 stability of the recovered color vector
• sample_words.py — frequency-random English-word sampling pool
• date_filter_test.py — Google Images date-range parameter (tbs=cdr:1,cd_min=…,cd_max=…) sanity check

Figure regeneration

• fig3_brand_similarity_transposed.py — Fig 3 (logo vs. web-image cosine, sorted leaderboard with full registered names)
• fig4_radar_only.py — Fig 4 (per-company 10-d radar comparison)
• regen_fig6.py — Fig 6 (Lowell − Johnson Δ with permutation-null SD error bars)
• make_serp_grids.py — moss / lilac schematic SERP grids used as components of Fig 1 and Fig 2

Data

• brand materials/brand_similarity_all_metrics.csv, cosine_summary_by_sector.csv — the logo-vs-web-image cosine similarities and per-sector means reported in the manuscript for the top-100 publicly traded companies (by market capitalization, TradingView, February 2025), used for Fig 3, Fig 4 and S4 Table
• csv/revision_results/ — per-analysis outputs
• brand materials/ — the brand list and numerical outputs for the top-100 companies. Google screenshots and corporate logo images are not redistributed (see brand materials/README.txt); they are re-scraped by LogoAndScreenshotOfBrands.ipynb
• Images/ — not redistributed (third-party Google Image content; see Images/README.txt); recreated locally by the scraping scripts
• external_compare/ — public datasets from prior approaches (Rathore UW-58 / BCP-37; comp-syn JzAzBz embeddings) used by validation scripts

Citation

Wang, S., & Park, J. Word–Color Association from Large-Scale Online Data: Methodology and Applications. PLOS ONE (under revision).