=Paper=
{{Paper
|id=Vol-3834/paper6
|storemode=property
|title=Quantitative Framework for Word-Color Association and Application to 20th Century Anglo-American Poetry
|pdfUrl=https://ceur-ws.org/Vol-3834/paper6.pdf
|volume=Vol-3834
|authors=Sungpil Wang,Juyong Park
|dblpUrl=https://dblp.org/rec/conf/chr/WangP24
}}
==Quantitative Framework for Word-Color Association and Application to 20th Century Anglo-American Poetry==
https://ceur-ws.org/Vol-3834/paper6.pdf
Quantitative Framework for Word-Color Association
and Application to 20th Century Anglo-American
Poetry
Sungpil Wang1 , Juyong Park1,∗
1
Graduate School of Culture Technology (GSCT), Korea Advanced Institute of Science & Technology (KAIST), Daejeon,
Republic of Korea
Abstract
Color symbolism is considered a critical element in art and literature, yet determining the relationship
between colors and words has remained largely subjective. This research presents a systematic method-
ology for quantifying the correlation between language and color. We utilize text-based image search,
optical character recognition (OCR), and advanced image processing techniques to establish a connec-
tion between words and their corresponding color distributions in the CIELch color space. We generate
a color dataset based on human cognition, and apply it for analysis of the literary works of poets asso-
ciated with Imagism and Black Arts Movements. This helps uncover the characteristic color patterns
and symbolic meanings of the movements with enhanced objectivity and reproducibility in literature
research. Our work has the potential to provide a powerful instrument for a systematic, quantitative ex-
amination of literary symbolism, filling in the gaps in prior analyses and facilitating novel investigations
of thematic aspects using color.
Keywords
Word Color Association, Digital Humanities, Distant Reading, Semantic Analysis, Lexical Relationship,
Lexical Discovery
1. Introduction
Color symbolism in art, literature, and anthropology refers to the use of colors as symbols
in various cultural contexts [13]. This effect is particularly pronounced in literature, where
colors assume roles that extend beyond being mere background elements. Researchers have
deeply investigated the application of colors in texts and their significance in literary works.
For instance, Andreeva [1] examines Stephen King’s ”Rose Madder” to show how color de-
picts characters’ emotional and psychological states, providing nuanced moral or philosoph-
ical implications. Mukhitdinovna [48] explores the linguistic aspects of color symbolism in
novels, highlighting how color characterizes objects, social attitudes, and moral concepts. This
indicates that colors convey complex meanings rather than simply serve as decorative ele-
ments. By choosing specific colors, authors can emphasize emotions or concepts related to
the story’s theme, making color an essential tool for literary expression. Additionally, color
reflects cultural, historical, and practical meanings [32]. Polshchykova and Polshchykova [52]
CHR 2024: Computational Humanities Research Conference, December 4–6, 2024, Aarhus, Denmark
∗
Corresponding author.
£ wangsp0317@kaist.ac.kr (S. Wang); juyongp@kaist.ac.kr (J. Park)
ȉ 0000-0003-4571-0017 (J. Park)
© 2024 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
589
CEUR
ceur-ws.org
Workshop ISSN 1613-0073
Proceedings
demonstrates how color in works from various races and cultures illustrates the diversity and
complexity of American culture, showing the colors’ different meanings in different contexts.
These studies explore the intricate circumstances surrounding the use of color, linking literary
works to the external environment. Furthermore, Underwood [66] reviews the growing impor-
tance of color terminology in literature from the 18th to the early 20th centuries, highlighting
how the trend has intensified over time. This evolution indicates a transition from theoreti-
cal descriptions to tangible depictions, with an emphasis on sensory details. It signifies the
increasing focus on sensory portrayal in writings as a means to mirror shifts in society and
culture. The heightened use of color underscores its ability to enhance narrative complexity
and emotional depth, making this aspect increasingly significant in modern literary analysis.
To summarize, color symbolism is an important literary technique that enriches the intricacy
of literary works, enabling readers to engage more profoundly with the text. This technique
has become more prominent in modern times; This amplifies the visual, emotional, cultural,
and societal significance of literary works, both inside the text itself and in a wider context.
Color representation in literary works is generally executed in two forms: Directly describ-
ing color using adjectives associated with certain hues [56, 67] or indirectly evoking a particu-
lar color through noun imagery [65, 33]. Researchers encounter two obstacles when assigning
color to specific phrases for analysis: First, term-image linkage is subjective. The researcher’s
common-sense perception, which depends on “objectivity and universality,” as expected by the
researcher, is frequently used to determine the associated colors of nouns [26, 25]. For instance,
a researcher might define a tree as green, expecting readers to accept this without question. But
there cannot be a one-to-one correlation between words and colors [6], and research findings
could be biased by participant assignments, casting doubt on objectivity and cross-cultural
generalization; Second, traditional humanities methods of the study of color symbolism have
shown consistent limitations. Prior research, which can be called ‘close reading’ studies, fre-
quently analyzes color based on a small number of terms and their subjectively assigned colors.
Critics point out that this approach, which seeks to comprehend the entire structure through
specific incidents, falls short of understanding the larger literary context [47]. Over-reliance
on a limited number of noteworthy color examples may impede a thorough comprehension of
the role of color in literature. Current color studies are restricted to counting direct color de-
pictions [56, 67], even when employing ‘remote reading’ techniques. Researchers still have to
categorize the images and colors that phrases represent, the subjectivity issue with matching
terms and pictures notwithstanding. However, one-to-one bonds between pictures and terms
are often deemed the most practical, though tagging all vocabularies is a daunting task due to
time requirements and the possibility of human errors or omissions [10, 9].
Through an intuitive use of massive imagery data, this study makes concrete the hitherto
ambiguous relationships between words and colors that were previously limited to the con-
ceptual realm. These data show the word-color associations that reflect real-world human
cognition and color perception, and the associations are quantified via generalized distribu-
tions in the color gamut. Words and pictures are correlated through a search platform, and
these images are translated into a color system aligned with human visual perception. The
color concept is then used to partition the color space and represent the picture as a distribu-
tion of familiar colors. These distributions are cross-examined on a general level to express
how strongly each color corresponds to the target term, providing a quantitative relationship
590
between words and colors that reflects human ideas and perceptions. By assigning colors to
words in a data-based objective manner, we provide a framework that identifies various col-
ors associated with words within a general frame of perception and expresses the relevance
quantitatively. This methodology is then applied to the complete works of Anglo-American
Imagist poet and Harlem Renaissance poet. Since the entire oeuvre published by a poet con-
stitutes their world of poetry, and an individual poem consists of the intentional arrangement
of terms, with each possibly expressing multiple colors, colors can be said to form a complex
relationship with poetry. Analyzing the multi-layered structure to discover patterns or groups
of colors used by the poet can quantitatively characterize the poet’s literary universe from the
perspective of color symbolism. Via this process, we can hope to conduct a distant reading
study on the colors in a large literary corpus and demonstrate its potential for new symbolism
studies.
2. Related Works
Recent years have witnessed the advancement of quantitative frameworks for modeling and
analyzing large-scale heterogeneous data, extending their scope to the field of liberal arts [46,
5, 17, 8]. Researchers are not only using traditional cultural study methodologies but also at-
tempting the application of statistical tools to cultural data [63]. These approaches enable a
comprehensive examination of extensive text and image data, along with intricate metadata, of-
fering additional parameters that complement personal readings of humanities resources with
objective and quantitative features [44, 36]. Technological progress has made it possible to
objectively depict universal human concepts, allowing for new methods to explore abstract
topics such as color symbolism. As a result, numerous efforts have been made to comprehend
the correlation between words and colors. These can be classified into three main types: First,
creating a word-color dataset by giving predetermined terms to participants in surveys and
collecting their replies on the corresponding colors. This endeavor has been motivated by the
limited availability of databases that match colors and phrases [22], occasionally utilizing gam-
ification elements to stimulate data gathering [35]; Second, using natural language processing
techniques to assign colors to words that are not directly color-related by analyzing the sur-
rounding terms tied to color expressions. Contextual reflection, aided by semantic embedding,
allows for the quantification of the relationship between text and color at a level that does
not involve images, using surrounding words to recognize color illustrations [50, 22]; Third,
although there are studies pairing terms and pictures using search engine results, they simply
use the RGB color space, forcing researchers to manually assign colors to visuals[50]. While all
these strategies use statistical procedures to allocate color compositions to words in the form
of ratios (weights) of colors, they do not specifically depict the relationship between the term
and the visuals it represents: they instead either use text or surveys, or directly assign images
to terms, with only a few people manually and subjectively partitioning and naming colors
within the color space. Some even restrict the composition of a word to an arbitrary number
of colors.
591
Figure 1: Word to Image to Color. The process of converting a word (example lilac) to color space
coordinates using Google Image Search and resizing. (a) The returned screen is cropped, resized to
10,000 pixels, then (b) the CIELch color space coordinates are extracted.
3. Methodology
We aim to establish a quantitative and set workflow linking words and colors, and then apply it
to textual data. A series of steps is necessary to assign color compositions to a word. We start
by obtaining images representing the given word and extracting the color values from each
pixel of the images. Next, the word can be said to be represented by the pixel values. Then we
compare the composition with that of general words to determine which colors are statistically
overrepresented that can then said to be characteristic of the given word. We later demonstrate
this methodology on a corpus of Anglo-American poets as an example of ‘objective’ research
in color and literature.
3.1. Word to CIELch List Conversion
Our goal is to represent a word using color values that reflect general human cognition more
accurately. To achieve this, we use Google Image search to concretize a word into multiple
images for enhanced statistical confidence and then convert the image pixels into the CIELch
color space, known for reflecting true human color cognition better than the simple RGB color
space. The procedure is illustrated in Figure 1.
3.1.1. Word to Image
We further combine the power of text-based image search with image analysis techniques. The
main idea of this approach is that the colors inherent in a concept are statistically predominant
in pictures indexed in relation to the word explaining that concept [50]. To utilize pictures
reflecting human expectations for any given word, we use Google’s image search. This modern
lookup engine provides access to illustrations uploaded and tagged by numerous users. We
designate one screen filled with images arranged in order of relevance to the word by the
592
algorithm as the representative image set for the word. The search, scrolling, counting images,
and taking screenshots are performed on the Selenium web browser automation library in
Python [62]. Subsequently, resizing is performed to represent a word as a set of 10,000 pixels.
As a result, an arbitrary word (search term) is represented by a screenshot uniformly resized
to 10,000 pixels.
Google Search Google’s search algorithm, which connects terms and pictures, evolved from
the PageRank system that evaluates webpage importance by analyzing link structure. If many
users access an image when searching for a specific word, the image is considered highly rel-
evant. Studies have used Google search to correlate language and universal illustrations, sup-
porting the relevance of visuals returned for specific texts [34, 69, 71]. Google’s search engine
has improved by integrating image information, user preferences, feature extraction, and text
annotation [68, 30], strengthening the relationship between input text and resulting images
[49, 29]. Therefore, using Google’s search engine is effective for linking terms and pictures.
Empirically, we found that some words either have few images or mostly text images. If fewer
than 30 pictures are found or more than three out of the first ten contain text, the word is as-
sumed to lack associated images. Pytesseract, a Python library with OCR capabilities, is used
to quickly recognize text within images [24].
Capturing Images When a word is probed, pictures typically related with that term are
displayed in order of relevance on one screen. We use screenshots to capture the benefits of al-
gorithms that apprehend human concepts while extending the scope of their universality. This
allows us to construct representative illustrations for words while capturing multiple highly
relevant images with minimal storage. All screenshots are taken after scrolling down 15% of
the screen to exclude website logos, search bars, and toolbars.
Resizing Images We use Python’s Pillow library to resize images [7]. Reducing the number
of pixels to be processed allows for a quicker link between words and colors while maintaining
color distribution. The resize method in the Pillow library uses High-LANCZOS and Antialias-
ing filters to effectively preserve the color distribution of the original image while minimizing
jaggies [11]. The LANCZOS filter, a resampling technique, maintains color distribution and
detail by handling high-frequency components, while the Antialiasing filter smooths the pixel
boundaries in the reduced image to minimize visual distortion. These two filters effectively
preserve the color distribution of the original image.
3.1.2. Images to CIELch Color List
The image’s pixel colors are converted from the RGB into the CIELch color space that takes into
account the features of human visual perception. In order to accurately measure the color com-
position of a term represented by the 10,000 pixels, it is necessary to know the precise location
of each pixel inside the color space. Our image processing involves utilizing the scikit-image
package to transform the RGB color values of every pixel into CIELch. The CIELch color space
is a polar coordinate representation of the CIELAB color space, which was devised by the Inter-
national Commission on Illumination (CIE) in 1976 with the intention of creating a color space
593
that is perceptually uniform [61, 57]. The color system defines colors in terms of chroma 𝑐∗
and hue ℎ∗ while preserving the lightness 𝐿∗ value of CIELAB. By transforming CIELAB color
space into polar coordinates, it becomes possible to comprehend and control colors in a more
instinctive manner. This approach highlights disparities in both hue and saturation, enabling
precise color comparison and analysis that aligns with human perception. This color space rep-
resents the entire range of human’s photopic (daylight) vision and provides a comprehensive
framework for color representation that closely matches how colors are perceived in real life.
Consequently, every pixel is located in the CIELch color space using a specific combination of
(𝐿, 𝑐, ℎ) values. Thus a word, via an image, translates to 10,000 (𝐿, 𝑐, ℎ) color space coordinates.
3.2. Removal of Grids
A resized screenshot contains images associated with the target word separated by white spaces
that need to be removed in principle, as they are not related to the visual representing the term.
The white spaces are different in each screenshot, and the detection is computation intensive,
making it impractical to perform it for every screenshot. For efÏciency therefore we calculate
the general total area of the white space in sample screenshots by choosing 100 random content
words and counting the average number of pixels with 𝐿 = 100 (the brightest white). We then
discount this number of white pixels from the screenshots.
Table 1
Basic Colors and Their Synonyms
Basic Color Synonyms
Red scarlet, vermilion, ruby, carmine
Orange tangerine, marmalade, orangish, apricot
Yellow yellowish, yellowy, gold, golden
Green greenish, verdant, leafy, greenery
Blue azure, cobalt, cerulean, ultramarine
Purple violet, purply, purplish, amethyst
Pink rosy, blushing, shellpink, rose
Black pitchblack, pitchdark, jetblack, blackish
Grey silver, slategray, smokegray, silvery
White snowywhite, milkwhite, milkywhite, chalkwhite
3.3. Establishing Color Standards
3.3.1. Selection of Basic Colors and Method of Color Composition
Basic Color Terms As the hue parameter of the colors of words, we use the 11 Fundamental
Color Terminologies that are generally employed across various civilizations [2]. The color of
a certain word is a composite blend of multiple hues, rather than a singular color. Thus, to
faithfully depict colors, we take into account both the hue and luminance of the visual that
represents the term, assigning chromatic and achromatic compositions to the word. Excluding
Brown, which traverses both chromatic and achromatic [54, 37], we select Red, Yellow, Blue,
594
Figure 2: The diagrams illustrate the merging frequency process for color distributions using syn-
onyms. (a) The combined frequencies show the hue distribution for blue-related words as an example
of chromatic colors. (b) The combined frequencies show the lightness distribution for white-related
words as an example of achromatic colors.
Orange, Green, Purple, Pink, White, Grey, and Black as representative colors. These represen-
tative colors include both Primary and Secondary Colors and have been typically recognized
as major colors in art and design since the 17th century [58, 14, 15].
Use of Color Synonyms To determine a more accurate and reliable color profile of the
named colors, we also consider their alternative names (synonyms) from the Oxford American
Writer’s Thesaurus (OAWT), creating a color standard that reflects broad cultural and linguis-
tic contexts. This standard is established by using the color values of collected color terms,
reflecting the universal expectations that representative colors hold as linguistic concepts. The
OAWT is an extensive and globally recognized reference, regularly updated with the most re-
cent evidence and research, making it a reliable benchmark [59, 55] with an easy accessibility
as a free component of the Apple Operating System. Using different linguistic expressions for
one color allows for a more realistic and precise setting of the universal range for that color.
By utilizing linguistic variations of the basic colors, we robustly set the categories for the ba-
595
sic colors while clearly reflecting the general concepts surrounding them. Including various
linguistic expressions for each color in the standard is a practical method for linking language
and color. Thus, we search for four synonyms for each core color term through the dictionary
and use them additionally to create specific color distinctions that correspond to people’s per-
ceptions. The basic color terms and their synonyms are listed in Table 1. The list of five words
representing one color is each converted to a CIELch list and merged into one, allowing for
duplicates. The combined basic color list is used to distinguish it from other colors as the color
value representing that color. For example, the basic color list for red is created by merging the
CIELch lists of “red,” “scarlet,” “vermilion,” “ruby,” and “carmine,” and it is distinguished from
the basic color lists of the other nine colors created in the same way. This process is illustrated
in Figure 2.
Chromatic and Achromatic Colors Three achromatic and seven chromatic hues make up
the ten fundamental color names. Our objective is to concurrently capture multiple different
hues that a word evokes, not merely a single hue. It makes it possible to express a phrase from
both chromatic and achromatic perspectives using the CIELch color space, which includes both
hue and lightness. Hue information links words to the seven chromatic colors, while lightness
information links to the three achromatic colors. We use seven basic color lists to divide the
hue into seven parts and three arrays to divide the lightness into three parts.
Chromatic Colors Hue identifies a color’s exact location on the color wheel, represented
numerically from 0.0∘ to 360.0∘ . We partition this plane into seven sections, assigning each
core color term to a section. The basic color list, merged from five arrays, is expressed as a
histogram for hue values. Each chromatic color is depicted as a distribution, showing its range
within the hue at a general perception level.
Achromatic Colors Lightness indicates how bright or dark a color is, expressed as a
value between 0.0 and 100.0 in the CIELch space. We divide this range into three sections for
the achromatic colors. Using histograms, we determine the range for each achromatic color
within lightness at a general perception level.
We apply a smoothing algorithm to the histograms to represent them as curved distri-
butions. The rolling mean helps mitigate noise and emphasize trends and continuity to ensure
the distribution represents the population. We use the rolling method from the Python library
pandas, calculating the average of ten data points at a time [64]. This process transforms the
histograms into smooth distributions. If multiple colors share a specific hue or lightness range,
the predominant color takes precedence in practice.
3.4. Color Allocation
Once the intervals for the basic colors are determined, we can represent a word with a CIELch
list and its color composition. As depicted in Figure 3, we use the ℎ values from the CIELch list
to express the word in terms of the proportions of the seven chromatic colors and the 𝐿 values
to express the proportions of the three achromatic colors. Each term can thus be described by a
596
Figure 3: Word to Color Ratio. The input word is converted into a CIELCh list and then transformed
into color ratios based on the assigned intervals from both chromatic and achromatic perspectives. The
resulting color ratios for lilac are listed in the table.
pair of chromatic and achromatic proportions, reflecting the visual diversity of the target term.
Additionally, by comparing these color compositions to those of general words, we can use
the 𝑍 -score to express which colors are emphasized in the target word, enhancing statistical
confidence in the word-color association.
3.4.1. Wildtype
After describing a word with colors and their proportions, it is crucial to determine how much
it differs from a typical, average word. This is similar to the biological notion of ”wildtype,”
which describes the mean genotype or phenotype [3]. This functions as the baseline state in
statistical analysis. Using the 𝑍 -score, we can analyze the tie between a word and its colors by
comparing the word’s color composition to the wildtype.
To create a reference wildtype dataset, we employ the WordNet interface from the NLTK
toolkit [4]. WordNet is a comprehensive database classifying semantic relationships between
words in English [45]. Using this approach, we randomly obtain a collection of 100 content
phrases with their accompanying photos. Each word is converted into CIELch lists and allo-
cated a proportionate composition for the 10 colors according to the agreed-upon color stan-
dards. And, we compute the 𝑍 -scores for the color composition ratios of each word, evaluating
how much a word highlights colors compared to the wildtype. However, discrepancies in the
distribution of colors can compromise the reliability of statistical analysis. To address this,
By utilizing the Yeo-Johnson transformation to standardize color composition ratios [70], we
enhance the precision of the Z-score calculation, resulting in a more reliable study of the re-
597
lationships between words and colors. This transformation can be performed using the ’Pow-
erTransformer’ module from the ’sklearn’ package [51]. This method decreases variation in
color composition and improves the accuracy of statistical analysis. By establishing a con-
nection between a random word and its visual representation, expressed as color values, and
building overarching color standards using color terminology and visual representations, we
examine the distribution of colors in a word and compare it to a reference group. This helps
ascertain whether it emphasizes specific colors more than average terms, providing a quantita-
tive demonstration of the associations between specific words and colors in universal human
perception.
3.5. Application
As an application of our method of word-color association, we have chosen to study Imagism
and the Harlem Renaissance, two prominent movements in Anglo-American literature. Imag-
ism, an early 20th-century literary movement, is known to emphasize the utilization of vibrant
and precise imagery and concise language, giving significance to visual and intuitive modes
of communication. In contrast, Harlem Renaissance utilizes written literature as a method
to emphasize African American identity and cultural confidence, effectively conveying social
and cultural messsages. This study seeks to evaluate the significance of color usages in the
literary realms of these movements and poets, showcasing a quantitative analysis of literary
works from a color-centric viewpoint. By employing this methodology, we hope to explore the
possibility of new directions in the study of color use.
3.5.1. Data Collection and Preprocessing
Table 2
Data Sources for Selected Authors
Movement Author Data Source Title
Google Digital Copy (2024) Legends (1921)
Internet Archive (2024) Can Grande’s Castle (1918)
Pictures of the Floating World (1919)
Imagism Amy Lowell
Project Gutenberg (2024) A Dome of Many-Coloured Glass (1912)
Sword Blades and Poppy Seed (1921)
Men, Women and Ghosts (1916)
Harlem Renaissance Georgia Douglas Johnson Internet Archive (2024) The Selected Works of Georgia Douglas Johnson (1997)
Our choice of Imagist poet is Amy Lowell, while our selected Harlem Renaissance poets is
Georgia Douglas Johnson. The works were compiled from their whole poetic world, or in in-
stances where the total collected works were not accessible, specific poetry collections were
combined. The collections were compiled from Project Gutenberg and freely distributed PDFs,
as summarized in the Table 2. We employed OCR (Optical Character Recognition) technology
[60], namely the image-to-text function of the Tesseract package [24], to extract tex from the
PDFs. Subsequently, we conducted data preparation on the text utilizing the pandas and NLTK
tools in Python [4]. The initial steps were eliminating stopwords, special characters, and nu-
598
merical values, and converting all uppercase letters to lowercase in the gathered text corpus
for each poet, followed by lemmatization to convert words to their base or root forms.
3.6. Network Construction and Analysis
The oeuvre of a poet is their creative universe, with each poem comprising words that evoke
various chromatic images. To understand the significance of colors in poetry, it can be helpful
to elucidate the connections among artistic aspects, language, and colors as a network structure.
By examining this structure and identifying patterns or clusters of colors used by the poet, we
gain a quantitative comprehension of their work in terms of color symbolism. This examination
reveals concealed significances and patterns in the poet’s use of colors.
Building Process We establish a tripartite network connecting the works, words, and colors
of each poet as separate node groups. A work is linked to every word it contains, and each word
is linked to corresponding colors based on the 𝑍 -scores. The relationships between works
and words are determined by frequencies, while ties between vocabularies and visuals follow
the aforementioned guidelines for assigning color compositions. This forms the basis of our
method for analyzing a poet’s works via color symbolism.
Analysis The principles for ascertaining color compositions were applied to every phrase in
the complete works of each poet. To identify the primary colors in the poet’s literary universe,
we made the projection to reduce the tripartite network into a bipartite network connecting
works and colors. Our objective was to assess the importance of colors and identify the most
significant ones in each poet’s oeuvre. By transforming the network, we characterize the poet’s
works from a color-oriented viewpoint and compute the centrality – network-based measure
of importance – of each hue.
In a tripartite network, we define artworks (titles of artworks) as 𝑃, words as 𝑊 , and colors
as 𝐶. The edges between 𝑃 and 𝑊 are determined by frequency, and the edges between 𝑊
and 𝐶 are weighted by 𝑍 -score. The process of projecting the artwork-word-color tripartite
network into an artwork-color bipartite network is as follows:
𝑤𝑖𝑗 = ∑ 𝑓𝑖𝑘 ⋅ 𝑧𝑘𝑗
𝑘∈𝑊
where:
• 𝑤𝑖𝑗 is the weight between artwork 𝑖 and color 𝑗.
• 𝑓𝑖𝑘 is the frequency between artwork 𝑖 and word 𝑘.
• 𝑧𝑘𝑗 is the Z-score between word 𝑘 and color 𝑗.
• 𝑘 ∈ 𝑊 denotes the sum over all words 𝑘 included in artwork 𝑖.
We apply the Birank random-walk-based algorithm to the resulting four works-color bi-
partite networks to calculate centrality [23]. This algorithm ranks nodes by considering the
network topology of both classes, minimizing information loss during successive projection
processes. Differences in color centrality between poets can be used to compare the influence
of colors chosen by poets in constructing their poetic worlds.
599
Figure 4: (a) Reference Group Color Distribution. The reference group (wild type) shows the mean
and standard deviation for each color ratio. Histograms display the frequency distribution of color
ratios for basic colors before applying the Yeo-Johnson transformation. (b) Z-Score Calculation for
the word lilac. Using the mean and standard deviation from the reference group (after Yeo-Johnson
transformation), the color ratios for the word lilac are transformed into Z-scores. Only colors with
Z-scores of 0.0 or higher are shown.
4. Results and Discussion
4.1. Color Standards and Intervals
As shown in Figure 5, the distribution of each of the seven chromatic colors is represented in
terms of hue, and the distribution of each of the three achromatic colors is represented in terms
of lightness. The area of each color distribution is smoothed based on the histograms of the
hue values of 50 000 pixels representing five words for each color, so the areas are equal. If
different distributions meet at a point, that point becomes the boundary separating the colors.
The finalized intervals are detailed in Table 3.
4.2. Wildtype
We selected 100 random content words and applied two procedures. The first is gap removal.
All images of the words are screenshots of small pictures representing the word, with small
gaps between them. We remove these gaps to emphasize the association between words and
colors. The wildtype is then used to determine which colors are overrepresented by a word
based on its color composition. We calculate the mean and standard deviation of the propor-
600
Figure 5: (a) Hue-Based Color Frequency Distribution. Frequency distribution of hue values for
seven chromatic colors, smoothed to show range and frequency. The color wheel summarizes the hue
ranges. (b) Lightness-Based Color Frequency Distribution. Frequency distribution of lightness
values for three achromatic colors, smoothed to reflect general perception. The bar chart represents
the ranges for white, grey, and black.
Table 3
Detailed Color Range Specifications
Basic Color Ranges
Red Hue (degree) : (19.3270, 45.6767)
Orange Hue (degree) : (45.6767, 80.9054)
Yellow Hue (degree) : (80.9054, 120.1811)
Green Hue (degree) : (120.1811, 203.2800)
Blue Hue (degree) : (203.2800, 293.8690)
Purple Hue (degree) : (293.8690, 331.2452)
Pink Hue (degree) : (331.2452, 19.3270)
Black Lightness (value) : (0.0000, 41.6539)
Grey Lightness (value) : (41.6539, 82.8975)
White Lightness (value) : (82.8975, 100.0000)
tions of each color in 100 screenshots. 𝑍 -scores can be computed based on these values. Sub-
sequently, the Yeo-Johnson transformation is applied, and Z-scores can be computed based on
the transformed data. The detailed results of these transformation parameters are presented in
the Table 4. These values serve as a baseline to help determine which colors are more prevalent
601
in a given word compared to others.
Table 4
Wildtype Statistics with Yeo-Johnson Transformation
Color Y-J Mean Y-J Std Lambda
Red 0.0351 0.0237 −11.3730
Orange 0.0685 0.0382 −6.1276
Yellow 0.1421 0.0721 −1.8217
Green 0.0925 0.0447 −5.1438
Blue 0.2963 0.1620 −0.0635
Purple 0.0243 0.0163 −16.7988
Pink 0.0154 0.0118 −24.7169
Black 0.1485 0.0609 −2.2434
Grey 0.5328 0.1621 +1.9217
White 0.2554 0.1101 −0.6978
White Grids The proportion of pixels with 𝐿 = 100.0 is determined for each random con-
tent word, and the mean is calculated. The proportion of pure white pixels is 0.1272. If the
proportion of pixels with 𝐿 = 100.0 in a specific word is less than this mean, all such pixels are
removed. Otherwise, only the equivalent proportion is removed.
Distribution and Statistics for Each Color Each random content word is converted into
a list of CIELch pairs. These lists are then expressed as proportions of colors according to
the established color standards. Thus, each color is represented by 100 such values of propor-
tion. As shown in Figure 4, the distributions and statistics differ by color, so we normalize the
distribution differences between color proportions and then calculate the mean and standard
deviation for each color. The transformed statistics are used to express the color composition
of the target word as a 𝑍 -score.
Application to Literary Works We examine the intricate connections of artworks, lan-
guage, and hues, and delineate associations between artworks and colors. This enables us
to identify the prominent hues associated with each poet, facilitating our understanding. To
evaluate the importance of colors for each poet, we utilize a threshold on the Z-scores, which
indicate the strength of the correlations between words and colors. Raising the threshold en-
tails implementing a more stringent criterion for the correlation between words and colors. As
the threshold rises, only colors with strong associations to words are retained in the analysis,
while those with weaker links are omitted. This method emphasizes the predominant hues in
a poet’s work, providing clearer insights into their creative decisions.
In Table 5, the prevalence of specific hues, especially black, intensifies as the threshold es-
calates for Georgia Douglas Johnson. At the initial threshold (0.000), where many colors are
evident, black has a prominent place. As the threshold escalates, black becomes progressively
central, and upon reaching a threshold of 2.576, black has a predominant birank value. This
602
Figure 6: The bipartite networks illustrate the connections between the works (outer circle) and colors
(inner circle) in the poetic works of Georgia Douglas Johnson and Amy Lowell. Nodes represent colors
and are sized by their birank-centrality, with a Z-score threshold of 1.282 indicating significant associ-
ations. And, The edges are represented as an MST (Minimum Spanning Tree). Both poets emphasized
red, purple and pink. (a) Georgia Douglas Johnson’s Poetic Networks. black is relatively prominent.
(b) Amy Loewll’s Poetic Networks. white is comparably emphasized, and the variation between the
colors is less pronounced.
use of color corresponds with the literary framework of the Black Arts Movement, noted for
its focus on Black identity and culture [27, 53, 16]. Amy Lowell exhibits a notably broad and
uniform application of color at the initial threshold (0.000). This equitable allocation of color
application corresponds with the tenets of the Imagism movement, which prioritizes tangi-
ble and varied imagery, encompassing both natural vistas and urban environments, to elicit
sensory experiences [21, 12, 20]. As the threshold rises, Lowell increasingly employs white,
underscoring a distinct pattern of color focus in contrast to Johnson.
5. Limitations and Future Works
The separation between chromatic and achromatic hues leads to the omission of brown, po-
tentially limiting detailed color analysis. Maintaining the CIELch space in a three-dimensional
form and assigning intervals could allow more differentiated colors.
Nevertheless, our color palette has the potential to accommodate the inclusion of new colors.
Each color is associated with the word and its synonyms, meaning new colors can be expressed
as distributions over the color range and occupy specific intervals. This flexibility allows re-
searchers to choose a variety of colors of interest to incorporate to the text being analyzed.
Currently, the network construction recognizes the significance of colors based on their asso-
603
ciations with artworks. Incorporating additional metrics like community detection algorithms
or topic modeling techniques can offer new insights into the influence of colors within text
networks and their contribution to thematic elements. Identifying clusters of subjects, phrases,
and colors can provide a comprehensive understanding of color symbolism in texts.
Acknowledgments
This work is supported by the KAIST Post-Al Research Grant, BK 21 FOUR Program, the Na-
tional Research Foundation of Korea (NRF-RS-2023-00245361, NRF-2020S1A5C2A03093177),
and the Culture, Sports, and Tourism R&D Program through the Korea Creative Content
Agency, funded by the Ministry of Culture, Sports, and Tourism in 2024 (KOCCA:RS-2023-
00270043, Contribution Rate: 50%)
References
[1] K. Andreeva. “The Impact of Colour Foregrounding on the Text of the Novel”. In: Inter-
national Journal of Linguistics 11.4 (2019), pp. 82–94.
[2] B. Berlin and P. Kay. Basic color terms: Their universality and evolution. Univ of California
Press, 1991.
[3] Biology Online. Wild Type - Definition from Biology Online Dictionary. https://www.biol
ogyonline.com/dictionary/wild-type. 2024.
[4] S. Bird, E. Klein, and E. Loper. Natural Language Processing with Python: Analyzing Text
with the Natural Language Toolkit. O’Reilly Media, Inc., 2009.
[5] J. Byszuk. “The Voices of Doctor Who–How Stylometry Can be Useful in Revealing New
Information About TV Series.” In: DHQ: Digital Humanities Quarterly 14.4 (2020).
[6] Y. Chen, J. Yang, Q. Pan, M. Vazirian, and S. Westland. “A method for exploring word-
colour associations”. In: Color Research & Application 45.1 (2020), pp. 85–94.
[7] A. Clark and Contributors. Pillow: The friendly PIL fork. Pillow. 2024.
[8] H. Craig and A. F. Kinney. Shakespeare, computers, and the mystery of authorship. Cam-
bridge University Press, 2009.
[9] S. K. Dandapat et al. “Tagging multi-label categories to points of interest from check-in
data”. In: IEEE Transactions on Emerging Topics in Computational Intelligence (2023).
[10] M. Diligenti, M. Gori, and M. Maggini. “Learning to tag text from rules and examples”. In:
AI*IA 2011: Artificial Intelligence Around Man and Beyond: XIIth International Conference
of the Italian Association for Artificial Intelligence, Palermo, Italy, September 15-17, 2011.
Proceedings 12. Springer. 2011, pp. 45–56.
[11] C. E. Duchon. “Lanczos filtering in one and two dimensions”. In: Journal of Applied Me-
teorology and Climatology 18.8 (1979), pp. 1016–1022.
[12] I. Fariha. “Poetic Image as Central Element in Early Imagist Poetry”. In: Journal of English
and Education (JEE) (2009), pp. 37–52.
604
[13] E. Feisner and R. Reed. “Color symbolism”. In: 2014, pp. 184–205. doi: 10.5040/97815013
03364.ch-014.
[14] J. Gage. Color and meaning: Art, science, and symbolism. Univ of California Press, 1999.
[15] J. Gage. Color and culture: Practice and meaning from antiquity to abstraction. Univ of
California Press, 1999.
[16] A. Gayle. “The Harlem renaissance: Towards a black aesthetic”. In: Midcontinent Ameri-
can Studies Journal 11.2 (1970), pp. 78–87.
[17] T. Gessey-Jones, C. Connaughton, R. Dunbar, R. Kenna, P. MacCarron, C. O’Conchobhair,
and J. Yose. “Narrative structure of A Song of Ice and Fire creates a fictional world with
realistic measures of social complexity”. In: Proceedings of the National Academy of Sci-
ences 117.46 (2020), pp. 28582–28588.
[18] Google Digital Copy. https://books.google.com. Google, 2024.
[19] Project Gutenberg. https://www.gutenberg.org. Project Gutenberg, 2024.
[20] B. S. Hama. “Imagism and Imagery in the Selected Poems of Major Imagist Poets”. In:
Koya University Journal of Humanities and Social Sciences 3.1 (2020), pp. 88–93.
[21] C. A. Hamilton. “Toward a cognitive rhetoric of imagism”. In: Style 38.4 (2004), pp. 468–
490.
[22] J. Harashima. “Japanese Word Color Associations with and without Contexts”. In: Pro-
ceedings of the Tenth International Conference on Language Resources and Evaluation
(LREC’16). 2016, pp. 2119–2123.
[23] X. He, M. Gao, M.-Y. Kan, and D. Wang. “Birank: Towards ranking on bipartite graphs”.
In: IEEE Transactions on Knowledge and Data Engineering 29.1 (2016), pp. 57–71.
[24] S. Hoffstaetter and Contributors. Pytesseract: A Python wrapper for Google Tesseract.
Pytesseract. 2024.
[25] D. Hunt. “Colour Symbolism in the Folk Literature of the Caucasus: TOPICS, NOTES
AND COMMENTS”. In: Folklore 117.3 (2006), pp. 329–338.
[26] J. Hutchings. “Folklore and symbolism of green”. In: Folklore 108.1-2 (1997), pp. 55–63.
[27] G. Hutchinson. The Harlem Renaissance in black and white. Harvard University Press,
1995.
[28] Internet Archive. https://archive.org. Internet Archive, 2024.
[29] T. Jiang and A.-H. Tan. “Discovering image-text associations for cross-media web in-
formation fusion”. In: European Conference on Principles of Data Mining and Knowledge
Discovery. Springer. 2006, pp. 561–568.
[30] Y. Jing and S. Baluja. “Pagerank for product image search”. In: Proceedings of the 17th
international conference on World Wide Web. 2008, pp. 307–316.
[31] G. D. Johnson. The selected works of Georgia Douglas Johnson. G.K. Hall, 1997.
[32] F. I. Kartashkova and L. E. Belyaeva. “Colour Meaning in English Literary Pieces”. In:
RUDN Journal of Language Studies, Semiotics and Semantics 13.1 (2022), pp. 201–212.
605
[33] M.-O. Kim. “A study on the color images of Jeong-Ju Seo Midang’s poetry - focusing
on the variation of desires”. Master’s thesis. Seoul: Seoul National University of Science,
Technology, Graduate School of Industry, and Engineering, 2012, pp. iv, 83.
[34] J. Kiros, W. Chan, and G. Hinton. “Illustrative language understanding: Large-scale vi-
sual grounding with image search”. In: Proceedings of the 56th Annual Meeting of the
Association for Computational Linguistics (Volume 1: Long Papers). 2018, pp. 922–933.
[35] M. Lafourcade, N. Le Brun, and V. Zampa. “Crowdsourcing word-color associations”. In:
Natural Language Processing and Information Systems: 19th International Conference on
Applications of Natural Language to Information Systems, NLDB 2014, Montpellier, France,
June 18-20, 2014. Proceedings 19. Springer. 2014, pp. 39–44.
[36] A. Lewis. “Modeling the humanities: Data lessons from the world of education”. In: In-
ternational Journal of Humanities and Arts Computing 10.1 (2016), pp. 51–62.
[37] D. T. Lindsey and A. M. Brown. “Sunlight and ”blue”: The prevalence of poor lexical color
discrimination within the ”grue” range”. In: Psychological Science 15.4 (2004), pp. 291–294.
[38] A. Lowell. A Dome of Many-Coloured Glass. ReadHowYouWant. com, 1912.
[39] A. Lowell. Can Grande’s castle. Macmillan, 1918.
[40] A. Lowell. Legends. Houghton MifÒin, 1921.
[41] A. Lowell. Men, women and ghosts. IndyPublish. com, 1916.
[42] A. Lowell. Pictures of the floating world. Macmillan, 1919.
[43] A. Lowell. Sword Blades and Poppy Seed. ReadHowYouWant. com, 1921.
[44] P. Martı́n-Rodilla, J. I. Panach, C. González-Pérez, and O. Pastor. “An experiment on ac-
curacy, efÏciency, productivity and researchers’ satisfaction in digital humanities data
analysis: dataset appendix”. In: (2016).
[45] G. A. Miller. “WordNet: a lexical database for English”. In: Communications of the ACM
38.11 (1995), pp. 39–41. doi: 10.1145/219717.219748.
[46] S. Min and J. Park. “Narrative as a Complex Network: A Study of Victor Hugo’s Les
Misérables”. In: Proceedings of HCI Korea. 2016, pp. 100–107.
[47] F. Moretti. Distant reading. Verso Books, 2013.
[48] K. I. Mukhitdinovna. “The Problem Of Symbolic Of Colours In Different Language
Cultures”. In: American Journal of Interdisciplinary Research and Development 4 (2022),
pp. 123–129.
[49] P. Nieuwenhuysen. “Information Discovery and Images A Case Study of Google Photos”.
In: 2018 5th International Symposium on Emerging Trends and Technologies in Libraries
and Information Services (ETTLIS). Ieee. 2018, pp. 16–21.
[50] G. Özbal, C. Strapparava, R. Mihalcea, and D. Pighin. “A comparison of unsupervised
methods to associate colors with words”. In: International Conference on Affective Com-
puting and Intelligent Interaction. Springer. 2011, pp. 42–51.
606
[51] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P.
Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher,
M. Perrot, and É. Duchesnay. “Scikit-learn: Machine Learning in Python”. In: Journal of
Machine Learning Research 12 (2011), pp. 2825–2830.
[52] O. Polshchykova and A. Polshchykova. “Colour signs-symbols in the artistic discourse
as components of the cultural semiosphere (on the material of the ethnic literature of the
20th-century USA)”. In: Filologičeskie nauki. Voprosy teorii i praktiki 16 (2023), pp. 1194–
1200. doi: 10.30853/phil20230202.
[53] R. J. Powell and D. A. Bailey. Rhapsodies in black: Art of the Harlem renaissance. Univ of
California Press, 1997.
[54] P. C. Quinn, J. L. Rosano, and B. R. Wooten. “Evidence that brown is not an elemental
color”. In: Perception & Psychophysics 43.2 (1988), pp. 156–164.
[55] I. G. Roberts, ed. The Oxford handbook of universal grammar. Oxford University Press,
2017.
[56] N. Romanyshyn. “Corpus Based Analysis of Color Names Function in Poetic Discourse”.
In: 2022 IEEE 17th International Conference on Computer Sciences and Information Tech-
nologies (CSIT). Ieee. 2022, pp. 111–114.
[57] J. Schanda. Colorimetry: understanding the CIE system. John Wiley & Sons, 2007.
[58] A. E. Shapiro. “Artists’ colors and Newton’s colors”. In: Isis 85.4 (1994), pp. 600–630.
[59] J. A. Simpson and E. S. C. Weiner, eds. Oxford English Dictionary. Vol. 3. Oxford University
Press, 1989.
[60] R. Smith. “An overview of the Tesseract OCR engine”. In: Ninth International Conference
on Document Analysis and Recognition (ICDAR 2007). Vol. 2. Ieee, 2007, pp. 629–633.
[61] C. Standard et al. “Colorimetry-part 4: CIE 1976 L*a*b*colour space”. In: International
Standard (2007), pp. 2019–06.
[62] S. Stewart and Contributors. Selenium WebDriver. Selenium. 2024.
[63] N. Tahmasebi, N. Hagen, D. Brodén, and M. Malm. “A Convergence of Methodologies:
Notes on Data-Intensive Humanities Research.” In: Dhn. 2019, pp. 437–449.
[64] T. pandas development team. pandas: Powerful Python data analysis toolkit. pandas. 2024.
[65] A. N. Toxirovna. “Colors and their artistic image creation features”. In: International Jour-
nal on Integrated Education 3.10 (2020), pp. 251–254.
[66] T. Underwood. Distant horizons: digital evidence and literary change. University of
Chicago Press, 2019.
[67] F. B. Wadsworth, J. Vasseur, and D. E. Damby. “Evolution of vocabulary in the poetry of
Sylvia Plath”. In: Digital Scholarship in the Humanities 32.3 (2017), pp. 660–671.
[68] Y. Wan, X. Liu, Y. Chen, et al. “Online image classifier learning for Google image search
improvement”. In: 2011 IEEE International Conference on Information and Automation.
Ieee. 2011, pp. 103–110.
607
[69] Y. Yao, J. Zhang, F. Shen, X. Hua, J. Xu, and Z. Tang. “Exploiting web images for dataset
construction: A domain robust approach”. In: IEEE Transactions on Multimedia 19.8
(2017), pp. 1771–1784.
[70] I.-K. Yeo and R. A. Johnson. “A new family of power transformations to improve normal-
ity or symmetry”. In: Biometrika 87.4 (2000), pp. 954–959.
[71] Y. Zhou, K. Chen, and X. Yang. “Google image search refinement: Finding text in images
using local features”. In: 2012 IEEE International Conference on Computer Science and Au-
tomation Engineering (CSAE). Vol. 1. Ieee. 2012, pp. 98–101.
608
Table 5
Color Centrality Ranking by Poet Against Different Thresholds
Imagism Harlem Renaissance
Amy Lowell Georgia Douglas Johnson
threshold : 0.000
color birank color birank
purple 0.136256 purple 0.156427
red 0.134740 red 0.134771
white 0.124037 pink 0.105000
blue 0.102301 white 0.104932
pink 0.099765 blue 0.102740
grey 0.098624 black 0.101768
orange 0.086580 grey 0.084472
yellow 0.080566 orange 0.084240
black 0.079782 yellow 0.071469
green 0.057347 green 0.054181
threshold : 1.282
color birank color birank
purple 0.159047 purple 0.192556
red 0.153338 red 0.159229
white 0.138381 pink 0.114978
pink 0.103135 black 0.112283
grey 0.090997 white 0.101010
blue 0.081868 orange 0.078249
orange 0.079721 grey 0.074453
black 0.079462 blue 0.069734
yellow 0.073915 yellow 0.063331
green 0.040136 green 0.034178
threshold : 1.645
color birank color birank
red 0.172395 red 0.186979
purple 0.161989 purple 0.183828
white 0.145447 black 0.124643
grey 0.094579 pink 0.116571
pink 0.092988 whit 0.104163
black 0.083757 orange 0.079978
orange 0.074665 grey 0.069749
yellow 0.070590 yellow 0.058651
blue 0.069440 blue 0.052452
green 0.034150 green 0.022987
threshold : 1.960
color birank color birank
red 0.202281 red 0.276942
white 0.160345 black 0.160740
purple 0.138880 pink 0.113437
grey 0.107884 purple 0.112160
black 0.090139 white 0.109497
yellow 0.087211 orange 0.078618
orange 0.074131 yellow 0.059716
pink 0.072172 grey 0.055331
blue 0.044089 blue 0.022949
green 0.022868 green 0.010608
threshold : 2.576
color birank color birank
white 0.316941 black 0.664419
yellow 0.238126 white 0.131132
black 0.226836 grey 0.116518
grey 0.217889 yellow 0.087931
orange 0.000207 - -
- - - -
- - - -
- - - -
- - 609- -
- - - -