💻 computer science

Cultural Targets, Structural Frames, Binding Morals: A Cross-Lingual Audit of Online Hate in Multicultural Singapore

This study of a 2025 Singaporean social media corpus reveals that while specific hate targets vary across English, Chinese, and Malay language communities, the underlying moral frameworks and threat structures of online hate remain remarkably consistent, suggesting that cross-lingual moderation strategies should prioritize shared moral and emotional patterns over culturally specific targets.

Original authors: Emilio Ferrara

Published 2026-06-23

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Emilio Ferrara

Original paper licensed under CC BY 4.0 (http://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

The Big Picture: A Three-Layer Cake of Hate

Imagine Singapore as a giant, multicultural potluck dinner. People are sitting at three different tables speaking English, Chinese, and Malay. They are all talking about the same group of "outsiders" (like immigrants, different religious groups, or people from mainland China), but they are doing it in their own languages.

The researchers wanted to know: Is the hate at these three tables totally different, or is there a hidden pattern?

They found that online hate is like a three-layer cake. The layers change as you go from the top to the bottom:

The Top Layer (The "Who"): This is culturally specific. It's like the menu. The English table is mostly complaining about Muslims and LGBTQ+ groups. The Chinese table is focused on the tension between local Chinese and people from mainland China. The Malay table is focused on religion and LGBTQ+ issues. What people hate depends entirely on which language table they are sitting at.
The Middle Layer (The "How"): This is shared. This is the recipe for the complaint. No matter the language, if someone hates a migrant worker, they frame it as an "economic threat" (stealing jobs). If they hate a religious group, they frame it as a "religious threat." The way they attack is surprisingly similar across all three languages.
The Bottom Layer (The "Why"): This is universal. This is the moral foundation. The researchers found that hate isn't usually about "fairness" (like "this is unfair"). Instead, it's almost always about purity (sanctity) and loyalty (betrayal). It's the feeling that "you are dirty" or "you are not one of us." This moral grammar is nearly identical in English, Chinese, and Malay.

The Main Takeaway: The target of the hate is local and specific, but the grammar (the structure and moral reasoning) of the hate is a shared language that everyone speaks, regardless of the words they use.

The "Amplifier" Problem: What Gets Heard vs. What Gets Said

The study also looked at what happens after someone posts a hateful comment. Do people just ignore it, or do they amplify it (like, share, or reply)?

Think of the internet as a giant sound system.

Production: This is how many people say something.
Resonance: This is how loud the sound system plays it back.

The researchers found a strange disconnect:

The Sound System is "Nativist": Even though people post hate against many different groups, the "sound system" (the algorithm and human engagement) only turns up the volume for anti-immigrant hate.
The Silence: Hate against religious groups or LGBTQ+ people gets posted, but it doesn't get amplified as much. It's like someone shouting into a microphone that is turned down.
The Result: What a community produces is not the same as what the community amplifies. The audience selectively boosts complaints about immigrants while letting other types of hate fade into the background.

The "Measuring Tape" Problem: Why We Can't Count Exact Numbers

One of the most important technical findings is about measurement.

Imagine trying to count how many "red" apples are in a basket using eight different color-blind robots.

The researchers asked eight different AI models to count "hate."
The Result: The robots couldn't agree. One robot might say a comment is hate; another might say it's just a strong opinion. Even the best robots only agreed with each other about 42% of the time on the trickiest cases.
The Lesson: Because the "measuring tape" (the AI) is shaky, the paper says we cannot trust the exact number of hate comments (e.g., "There are exactly 5,000 hate comments").
The Solution: Instead of counting the total amount, the researchers focused on the structure. They looked at the patterns (e.g., "Hate against Group A is always framed as an economic threat"). These patterns were clear and consistent, even if the exact counts were fuzzy.

The "Event" Test: Does Hate Spike During Big News?

Finally, the researchers checked if hate exploded during big 2025 events in Singapore, like elections, arrests, or religious tensions.

The Expectation: You might think that when a big news story breaks, hate comments would spike like a wave.
The Reality: The hate volume was a steady, flat line. It didn't jump up during elections or arrests.
The Analogy: It's like a river that flows at a constant speed. A big rock (a news event) thrown in might make a splash, but it doesn't change the speed of the river. The hate is a constant background noise, not a reaction to specific news cycles.

Summary of Key Findings

Targets are local, but the "grammar" of hate is global. (Who they hate changes by language; how and why they hate stays the same).
Hate is driven by "purity" and "loyalty," not fairness.
The audience is selective. They amplify hate against immigrants but ignore hate against other groups.
We can't count hate perfectly. AI tools disagree too much to give an exact number, so we must focus on the patterns of hate instead.
Hate is steady, not reactive. It doesn't spike just because of big news events.

Technical Summary: Cultural Targets, Structural Frames, Binding Morals

Problem Statement
Online hate research has traditionally focused on single-language, single-platform contexts, primarily addressing detection from the sender's perspective. Two critical questions remain under-explored: (1) whether the structural components of hate (targets, threat frames, moral foundations) are shared across co-existing language publics within a single society, or if they are culturally idiosyncratic; and (2) whether the hate a community produces is the same hate its audience amplifies. Furthermore, the field faces a methodological challenge: Large Language Models (LLMs) used as annotators often disagree on absolute hate prevalence, making cross-lingual comparisons of raw rates unreliable.

Methodology
The study utilizes a Singapore-centric corpus from 2025, comprising 31.0 million items across Facebook, Reddit, and YouTube. The dataset includes 1.76 million comments mentioning eleven Singapore-salient out-groups (e.g., PRC nationals, migrant workers, religious groups, LGBTQ) in English, Simplified Chinese, and Malay.

Annotation Pipeline: The authors employ a two-stage classification process. Stage 1 uses a high-recall lexical and prototype filter. Stage 2 employs an isolated binary hate decision using LLMs, with explicit null hypotheses (e.g., neutral mentions, policy criticism) to avoid inflating prevalence. Confirmation requires consensus across two independently worded prompts.
Model Benchmarking: Eight open-source LLMs were benchmarked against a 211-item human-adjudicated "hard" gold set. Agreement ceilings were found to be low ( $\kappa \approx 0.42$ ), indicating that absolute prevalence is ill-defined at the annotator level. Consequently, the study adopts Phi-4 (14B parameters) as the primary labeler due to its superior balance of accuracy (0.95), Cohen's $\kappa$ (0.91), and recall (1.00) on an independent manual check. All findings are replicated using a second independent model (Qwen2.5) to ensure robustness.
Analysis: The analysis focuses on relative structural comparisons rather than absolute rates. It measures cross-lingual divergence using Cramér's $V$ and Total Variation distance across four axes: Target Salience, Threat Frame, Moral Foundation, and Emotion. Statistical significance is tested using $\chi^2$ tests with Benjamini–Hochberg FDR control. Resonance (amplification) is modeled via within-thread paired comparisons on Facebook, controlling for thread, length, and time.

Key Contributions

Layered Cultural Contingency: The paper proposes a falsifiable thesis that online hate is structured by cultural contingency. It demonstrates that cross-lingual divergence decreases monotonically as one moves from what is hated (targets) to how it is framed (threats) to why it is hated (moral foundations).
Production-Resonance Dissociation: The study foregrounds the reception side of hate, showing that audience amplification is selectively nativist and does not linearly track production volumes.
Measurement Validity: It establishes that for contested constructs like hate, absolute prevalence is not well-defined across LLM annotators (ceilinging at $\kappa \approx 0.42$ ). The paper argues for reporting relative structure as the primary unit of analysis, with absolute rates treated only as calibrated ranges.

Results

Targets (Culturally Specific): Target salience varies significantly by language ( $V=0.25$ ). English discourse is broad (focusing on Muslims, PRC, LGBTQ); Chinese discourse centers on the local-Chinese/PRC axis; Malay discourse focuses on religion and LGBTQ.
Threat Frames (Structural): While targets differ, the mapping between targets and threat frames is robust. Religious groups are framed via religious threats; migrants/foreign nationals via economic, safety, or replacement threats; and local-Chinese via political power. Anti-Muslim hostility shows the highest cross-lingual invariance.
Moral Foundations (Near-Universal): The moral grammar of hate is highly shared across languages ( $V=0.08$ ). Hate is predominantly driven by binding moral foundations (Sanctity/55–75%, Loyalty, Authority) rather than individualizing foundations (Fairness).
Expression: Hate is contempt-driven (94–96% across languages) and angry, with frequent dehumanization (21–35%). It is primarily an out-group grievance (anti-immigration) rather than an anti-system one; anti-immigration and anti-establishment stances are negatively associated.
Resonance: Hateful comments are generally amplified less than neutral mentions. However, amplification is selectively nativist: anti-CECA, anti-PRC, and anti-migrant-worker hate are preferentially amplified, while hate against religious and LGBTQ groups is suppressed.
Event Reactivity: Hate volume does not spike significantly around major 2025 Singapore events (e.g., elections, racial incidents), suggesting a steady baseline rather than event-driven volatility.

Significance and Claims
The paper claims that treating "online hate" as a single, monolithic prevalence obscures a layered structure. The central finding is that while the targets of hate are locally selected and culturally plastic, the grammar of hate (the threat frames and binding moral foundations) is cross-lingually shared.

This has direct implications for cross-lingual content moderation:

Policy Transfer: Detectors or policies keyed on threat frames are more likely to transfer across languages than those keyed on specific targets, which require localization.
Moderation Goals: Reducing the production of hate is distinct from reducing its reach. Since amplification is selectively nativist, moderation strategies must account for the fact that the audience filters hate differently than it is produced.
Methodological Rigor: The study cautions against relying on absolute hate prevalence metrics from LLMs alone, advocating for relative structural analysis and human-anchored calibration.

The authors maintain a modest scope, noting that their findings are specific to the Singapore context (one society, one year) and that the corpus is actor-less (analyzing comments, not users). They emphasize that their theoretical lenses are used descriptively, not causally.