💬 NLP

SoS: Analysis of Surface over Semantics in Multilingual Text-To-Image Generation

This paper introduces the first comprehensive analysis of the "Surface-over-Semantics" (SoS) phenomenon in multilingual text-to-image models, revealing that most models prioritize surface-level linguistic cues over semantic meaning, leading to culturally stereotypical depictions across various languages and cultures.

Original authors: Carolin Holtermann, Florian Schneider, Anne Lauscher

Published 2026-01-26

📖 4 min read☕ Coffee break read

CC0 1.0

Original authors: Carolin Holtermann, Florian Schneider, Anne Lauscher

Original paper dedicated to the public domain under CC0 1.0 (http://creativecommons.org/publicdomain/zero/1.0/). ✨ 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

Imagine you have a magical art machine that draws pictures based on what you type. If you tell it, "Draw a person from Germany," it should draw a German person. But what happens if you type the exact same request in a different language, like "Zeichne eine Person aus Deutschland" (German) or "Dibuja una persona de Alemania" (Spanish)?

This paper investigates a strange glitch in these machines called Surface-over-Semantics (SoS).

The Core Problem: The "Accent" vs. The "Meaning"

Think of a text-to-image prompt like a letter sent to an artist.

The Semantics (The Meaning): The actual content of the letter. "I want a picture of a German person."
The Surface (The Look): The language and script the letter is written in. Is it written in English, Hindi, or Chinese?

The researchers found that many of these AI art machines are like bad translators who focus on the accent rather than the story. When you ask for a "German person" in English, the machine draws a German person. But if you ask for the same "German person" using a different language (like Hindi or Finnish), the machine often ignores the "German" part and instead draws something that looks like a stereotype of the language you used.

It's as if you asked a chef for "Spaghetti" in Italian, and instead of making pasta, they served you a pizza because they heard the word "Italian" and assumed that's what you wanted, ignoring the word "Spaghetti."

The Experiment: A Global Taste Test

To prove this, the researchers created a massive "taste test":

The Menu: They picked 171 different cultural identities (like "Japanese," "Nigerian," "Brazilian").
The Languages: They translated the request "A photo of a [Culture] person" into 14 different languages (including English, Arabic, Chinese, Finnish, etc.).
The Artists: They asked 7 different AI art machines to draw these pictures.
The Scorecard: They invented a new way to measure the results, called the SoS Score.

The SoS Score is like a "Bias Meter":

If the score is positive, the machine listened to the meaning (e.g., It drew a German person when asked for a German person, regardless of the language).
If the score is negative, the machine listened to the surface (e.g., It drew a stereotypical "Finnish" scene when asked for a German person, just because the prompt was in Finnish).

What They Found

Most Machines are "Surface-Heavy": Out of the 7 machines tested, 6 of them were heavily influenced by the language used. If you typed in a language the machine wasn't very good at, it would often default to drawing stereotypes associated with that language's script or culture, rather than the culture you actually asked for.
The "Deep Dive" Effect: The researchers looked inside the machines' "brains" (specifically the layers where the text is processed). They found that as the message travels deeper into the machine, the bias gets stronger. It's like a game of "Telephone" where the message gets more distorted the further it goes; by the time the machine starts drawing, the language's "accent" has completely drowned out the original meaning.
Visual Stereotypes: When the machines got confused, they didn't just draw random things; they drew cultural stereotypes.
- Prompts in Finnish often resulted in snowy forests.
- Prompts in Amharic (an Ethiopian language) often resulted in dry, sandy deserts.
- Prompts in Chinese or Japanese sometimes included scary elements like blood or scars, which had nothing to do with the actual request.

The One Exception

One machine, called AltDiffusion, was the "good student." It mostly ignored the language accent and stuck to the meaning, drawing the requested culture regardless of the language used. However, even this machine struggled a bit with languages it hadn't seen much during training.

Why This Matters

The paper concludes that while these AI tools are getting better, they still have a blind spot. They are too easily swayed by the look of the words (the language) rather than the meaning of the words. This means that if you use these tools in your native language, you might get a picture that reflects a stereotype of that language rather than the specific person or thing you asked for.

The researchers suggest that to fix this, we need to train these machines to pay more attention to the "story" (semantics) and less to the "accent" (surface), especially for languages that aren't English. They also propose their "Bias Meter" (SoS Score) as a tool to help developers check if their machines are being fair.

Technical Summary: Analysis of Surface-over-Semantics in Multilingual Text-To-Image Generation

Problem Statement

Text-to-image (T2I) models are increasingly utilized globally, yet they exhibit high sensitivity to input languages. While semantically equivalent prompts (e.g., a prompt and its translation) should theoretically yield similar outputs, prior research indicates that T2I models often produce culturally stereotypical depictions when faced with non-English languages. This phenomenon suggests a tension between the surface form of the input (the specific language or script) and the semantic content (the actual description of the subject). Previous studies have identified this issue but have been limited by small-scale analyses covering only a handful of languages and cultures, and crucially, they lack a comprehensive method to quantify the degree to which models prioritize surface cues over semantic meaning. This gap hinders the development of globally inclusive and culturally fair T2I models.

Methodology

Dataset Construction

The authors constructed a novel multilingual dataset to systematically analyze this behavior:

Cultural Scope: 171 unique cultural identity groups were selected from the SeeGULL dataset, refined to remove duplicates and standardize names.
Linguistic Scope: Prompts were translated into 14 languages (English plus 13 others), selected to ensure typological diversity (covering Indo-European, Afro-Asiatic, Sino-Tibetan, Turkic, Koreanic, Japonic, and Uralic families) and model coverage.
Prompt Templates: Three base templates were defined (e.g., "A photo of a {ci} {pj}"), instantiated with cultural identities and person terms ("person," "woman," "man"), resulting in 1,539 unique English prompts.
Translation: Native speakers fluent in both English and the target languages translated the prompts, ensuring distinctions between gendered and gender-neutral terms were preserved.
Image Generation: Seven state-of-the-art T2I models were evaluated: AltDiffusion-m9 (AD), Kandinsky-2-1 (K21), Kandinsky-3 (K3), Stable Diffusion v2.1 (SD21), Stable Diffusion XL (SDXL), Stable Diffusion 3 (SD3), and FLUX.1-dev (FX). Images were generated using a fixed random seed (42).

Quantification: The Surface-over-Semantics (SoS) Score

To address the lack of a suitable metric, the authors introduced the SoS Score, an embedding-based measure designed to quantify the tension between surface and semantics without relying on textual comparisons that may be biased by language representation quality.

Reference Vectors: Two reference vectors are computed by pooling images across all models:
- $\vec{surl}$ : The average embedding of all images generated from prompts in a specific language $l$ (representing surface form).
- $\vec{semc}$ : The average embedding of all images generated for a specific cultural identity $c$ (representing semantic content).
Calculation: For an output image $o_{c,l}$ , the score is the difference between its cosine similarity to the semantic vector and the surface vector:
$SoS_{c,l}(o_{c,l}) = \cos(\vec{semc}, \vec{e_o}) - \cos(\vec{surl}, \vec{e_o})$
Interpretation: Scores range from $[-1, 1]$ . Negative values indicate a stronger alignment with the surface form (language), while positive values indicate alignment with the semantic content.

Validation and Analysis

Human Validation: A subset of images was annotated by humans to verify if the image matched the cultural identity (semantics) or the language-associated culture (surface). The SoS score achieved 74.0% accuracy against human judgments, with higher precision (94.8%) in detecting surface-level tendencies compared to CLIPScore.
Layer-wise Analysis: Using DiffusionLens, the authors analyzed how SoS tendencies evolve across the layers of text encoders in SD21 and K3.
Visual Manifestation: The study employed Visual Question Answering (VQA) using Qwen2-VL-7B-Instruct to generate detailed image descriptions. These descriptions were analyzed using the "Fighting Words" method to identify tokens distinctive to specific language-model pairs. These terms were then compared against known cultural stereotypes from the SeeGULL dataset.

Key Results

Model and Language Variations

Prevalence of Surface Bias: All models except AltDiffusion (AD) exhibited strong surface-level tendencies in at least two languages. AD generally showed semantic tendencies, though it still leaned toward surface cues for Hindi, Amharic, and Finnish.
Model Differences: FLUX (FX) displayed strong negative SoS scores (surface bias) for Korean, Finnish, Japanese, and Chinese. Kandinsky-3 (K3) showed the smallest negative mean score (closest to neutrality), while SD3 showed a wide range of scores, suggesting an amplification of stereotypes in both directions.
Layer Evolution: In later layers of the text encoders, the bias toward the prompt's surface form became more pronounced. For K3, European languages shifted toward neutrality in later layers, whereas Asian languages moved toward more negative (surface-biased) scores.

Correlations and Generalizability

Script Similarity: Pearson correlations of SoS scores revealed strong similarities among Indo-European languages using Latin scripts (e.g., Italian and Spanish showed 0.7 correlation). However, script similarity did not strictly dictate tendencies; for instance, Chinese and Japanese showed low correlation despite script overlap, while Chinese and Korean showed high correlation.
Concept Robustness: The SoS tendencies observed for "person" prompts were highly correlated (Pearson $r = 0.876$ ) with those observed for "house" prompts, indicating the metric generalizes beyond human subjects.

Visual Manifestations of Bias

Stereotype Coverage: Languages with strong surface tendencies often triggered culturally stereotypical visual depictions. For example, FLUX generations for Chinese prompts covered 61.9% of Chinese visual stereotypes, while Finnish prompts covered none (due to limited stereotype data for Finnish in the reference set).
Specific Associations: VQA analysis revealed systematic patterns:
- Finnish: Frequent generation of snowy forest scenes.
- Amharic: High frequency of terms like "arid," "sandy," and "bush."
- Chinese/Korean/Japanese/Arabic: Alarmingly, a significant portion of images included terms like "blood," "scars," or "splatter," raising safety concerns.
- Hindi/Amharic/Asian Languages: Frequent association with "butterflies," "fantastical," and "dreamlike" aesthetics, suggesting unfamiliar languages are linked to surreal imagery.
- Hindi: Despite AD's overall semantic tendency, images prompted in Hindi frequently included "saree" and "Indian," indicating persistent cultural associations.

Significance and Claims

The paper claims to provide the first systematic analysis of T2I models' "Surface-over-Semantics" (SoS) tendencies across a broad range of cultures and languages. Its primary contributions are:

A Novel Metric: The introduction of the SoS score, which quantifies the tension between input language and semantic meaning using only generated image embeddings, thereby reducing reliance on potentially biased textual reference models like CLIP.
Empirical Evidence: Demonstrating that the majority of tested models prioritize surface forms over semantics in multiple languages, a tendency that intensifies in the deeper layers of text encoders.
Visual Correlation: Establishing a direct link between high surface-level tendencies and the generation of culturally stereotypical or potentially harmful visual content (e.g., violence, specific aesthetic tropes).
Contextual Nuance: The authors argue that the "desirable" SoS score is context-dependent; while semantic alignment is crucial for general concepts, language-specific grounding may be appropriate for culturally specific prompts. They advocate for future systems to dynamically adapt based on prompt type.

The authors conclude that while SoS scores can serve as a metric to guide model training toward fairer outputs, they are not a comprehensive fairness measure and must be coupled with other evaluations of generation quality and factual accuracy.