Visual–semantic coherence of Chinese colour metaphor lexemes: What interpretable visual features reveal about multimodal meaning stability
This study demonstrates that the visual–semantic coherence of Chinese colour metaphor lexemes is significantly predicted by interpretable visual features—particularly horizontal symmetry and consistent visual variability—rather than just central image tendencies, revealing that multimodal meaning stability relies on the recognizability of visual patterns across diverse digital instantiations.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
In the digital age, words rarely travel alone. When we search for a term online, we are rarely met with text alone; we are greeted by a flood of images that our brains instantly try to match with the meaning of the word. This pairing of language and vision is a fundamental part of how we understand the world, but it is not always a perfect match. Some words, like "apple" or "mountain," bring to mind a very clear, consistent picture. Others, especially those used in complex cultural or metaphorical ways, can summon a chaotic mix of images that seem to have little in common. Researchers have long known that concrete objects are easier to picture than abstract ideas, but a new question has emerged: when a word does have a visual connection, what makes that connection strong or weak? Is it the meaning of the word itself, or is it something about the pictures we find when we look for it?
A team of researchers set out to answer this by examining Chinese colour words used in metaphorical ways. In Chinese, colour terms like "red" or "black" are not just about light wavelengths; they carry heavy cultural weight. "Red" can signify revolution or good fortune, while "black" might imply something illegal or illicit. These words form a rich network of compounds, such as "red song" for revolutionary music or "black market" for illegal trade. The researchers wanted to know if these metaphorical words have a stable visual identity in the modern digital world. Do the images that appear when you search for them actually look like what the words mean, or is the connection broken? To find out, they built a massive dataset of 184 such colour-based words. For each word, they gathered 100 images from a major Chinese search engine and 100 examples of the word used in sentences from a large collection of digital text.
The core of their investigation was to measure how well the visual meaning of a word matched its textual meaning. They used a sophisticated computer system trained to understand both images and language, which could assign a score indicating how closely the "idea" of the images aligned with the "idea" of the text. If the score was high, the word had a stable visual-semantic connection; if it was low, the images and text were drifting apart. But the researchers did not stop at just measuring the score. They wanted to understand why some words scored well and others did not. They extracted dozens of specific, measurable properties from the images themselves—things like how symmetrical the pictures were, how much colour varied across the set, how busy or cluttered the images looked, and where the eye was naturally drawn within the frame. They then asked if these visual properties could predict the alignment score.
The results revealed a surprising truth about how meaning is constructed in the digital age. The researchers found that the visual properties of the images explained a significant portion of the variation in how well the words and pictures matched. In fact, these human-interpretable visual features were better at explaining the connection than the internal, complex mathematical representations of the computer model itself. The most powerful predictor was not what the images typically looked like on average, but how consistent they were. A word achieved a strong connection when the images associated with it shared a recognizable visual pattern, even if that pattern was simple. For instance, words whose images showed a high degree of horizontal symmetry or a balanced distribution of visual attention tended to have much stronger connections between their text and image meanings.
Conversely, the study showed that a lack of visual consistency was a major barrier to meaning. When the images for a single word were wildly different from one another—some symmetrical, some chaotic, some bright, some dark—the connection between the word and its visual meaning weakened, regardless of how clear the word's definition was. This finding challenges the old idea that meaning is just a fixed picture in our heads. Instead, it suggests that meaning is a dynamic stability. A word holds its visual meaning together not because every single image is identical, but because the collection of images evokes a coherent, recognizable visual experience. The study also ruled out the idea that the computer model's own internal "black box" features were the best explanation for this stability; the simple, observable traits of the images, like symmetry and colour balance, told a clearer story.
Ultimately, this work provides a concrete map of how language and vision interact in the modern world. It shows that for metaphorical words to maintain a stable meaning in the digital environment, the images they attract must be visually consistent. If the visual world surrounding a word becomes too chaotic or inconsistent, the word's meaning begins to fracture, even if the text remains the same. The researchers demonstrated that we can understand this complex relationship by looking at the basic building blocks of the images themselves. The stability of a word's meaning in the digital age depends less on the word's definition and more on the visual harmony of the world it inhabits.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.