💬 NLP

Cross-lingual Offensive Language Detection: A Systematic Review of Datasets, Transfer Approaches and Challenges

This paper presents the first systematic review of cross-lingual transfer learning techniques for offensive language detection in social media, analyzing 67 studies to categorize datasets, summarize transfer strategies, and identify current challenges and future research directions.

Original authors: Aiqi Jiang, Arkaitz Zubiaga

Published 2026-04-02

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Aiqi Jiang, Arkaitz Zubiaga

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

Imagine the internet as a giant, bustling global town square. In this square, people from every corner of the world are chatting, sharing jokes, and arguing. But like any crowded place, there are also bullies, trolls, and people shouting hate speech.

The problem is that the "town guards" (the AI systems designed to catch these bullies) are mostly trained to speak only one language: English. They are great at spotting a rude comment in English, but if a bully switches to Spanish, Hindi, or Arabic, the guards are often confused and let the bad behavior slide.

This paper is a comprehensive guidebook written by two researchers, Aiqi Jiang and Arkaitz Zubiaga, to help us build better guards who can speak many languages. They looked at 67 different studies to figure out the best ways to teach an AI to understand offensive language across different cultures.

Here is the breakdown of their findings, using some simple analogies:

1. The Core Problem: The "Language Barrier"

Imagine you have a master chef who knows how to cook a perfect spicy curry (English). You want to teach a new apprentice how to cook a spicy stew in a different country (a low-resource language like Swahili or Bengali), but you have no recipes or ingredients for that specific stew.

The researchers asked: How do we take the chef's knowledge of "spiciness" and apply it to a completely different cuisine without starting from scratch?

2. The Three Ways to Teach the AI (The Transfer Methods)

The paper categorizes the solutions into three main "teaching styles," which they call Transfer Learning. Think of these as three different ways to help the apprentice learn:

Instance Transfer (The "Translator" Method):
- The Analogy: You take the English recipe, translate it word-for-word into the new language, and give it to the apprentice.
- How it works: The AI takes a hateful English sentence, translates it into the target language, and says, "Okay, this translated sentence is also hate speech." It's like using a dictionary to copy-paste labels.
- Pros/Cons: It's fast, but if the translation is bad (or if a joke doesn't translate well), the AI gets confused.
Feature Transfer (The "Shared Vocabulary" Method):
- The Analogy: Instead of translating the whole recipe, you teach the apprentice that "spicy" and "hot" mean the same thing in both English and the new language. You build a shared mental map where similar concepts sit next to each other, regardless of the language.
- How it works: The AI learns that the feeling of an insult is similar across languages, even if the words are different. It uses "multilingual embeddings" (a fancy word for a shared dictionary of meanings) to connect the dots.
- Pros/Cons: Great when you don't have exact translations, but it can miss subtle cultural nuances.
Parameter Transfer (The "Brain Upgrade" Method):
- The Analogy: You take the master chef's brain (the pre-trained AI model) and give it to the apprentice. The apprentice already knows how to think about food; they just need a little bit of practice with the local ingredients.
- How it works: This is the most popular method today. Researchers use massive AI models (like mBERT or XLM-R) that have already "read" the internet in dozens of languages. They then "fine-tune" this brain with a tiny bit of data from the new language.
- Pros/Cons: This is usually the most powerful method, but it requires a lot of computing power.

3. The Hurdles (Why It's Still Hard)

Even with these smart methods, the researchers found several "potholes" on the road:

The "Cultural Blind Spot": What is considered an insult in one culture might be a compliment in another. For example, calling someone a "fox" might be an insult in English (sly) but a term of endearment in another culture. The AI often misses these cultural subtleties.
The "Code-Mixing" Chaos: People on social media often mix languages in one sentence (e.g., "Hinglish" is a mix of Hindi and English). It's like someone speaking English but suddenly throwing in a French phrase. The AI gets dizzy trying to figure out which language rules to apply.
The "Data Desert": For many languages, there is almost no data to train on. It's like trying to teach someone to drive in a country where no one has ever built a road.
The "Translation Trap": Sometimes, machine translation kills the meaning. A sarcastic joke might translate literally and lose its sting, making the AI think it's harmless when it's actually toxic.

4. The Future: What's Next?

The researchers suggest a few exciting directions for the future:

Don't rely on "Zero-Shot" alone: You can't just throw a model at a new language and expect it to work perfectly. You need at least a little bit of local data (a few hundred examples) to "fine-tune" the system.
Use "LLMs" (Large Language Models): New super-smart AI models (like GPT-4) are getting really good at understanding context without needing massive training. They might be the key to solving the "cultural blind spot."
Human Help: We need more diverse human annotators (people from the specific cultures) to help label the data. An AI can't understand a cultural nuance unless a human from that culture explains it.

The Bottom Line

This paper is a map for the future of internet safety. It tells us that while we have made huge progress in teaching AI to speak many languages, we still need to be careful about culture, context, and data quality. We can't just translate the rules; we have to understand the people behind the words.

The authors have also made a "toolkit" available online, listing all the datasets and methods they found, so other researchers can pick up where they left off and build a safer, more inclusive internet for everyone.

1. Problem Statement

The proliferation of offensive language and hate speech (HS) on social media platforms poses significant societal risks. While Natural Language Processing (NLP) has made strides in detecting such content, most research focuses on monolingual settings (primarily English). However, social media is inherently multilingual.

The Core Challenge: Developing effective detection models for low-resource languages (languages with limited annotated data) is difficult due to data scarcity, cultural nuances, and linguistic diversity.
The Proposed Solution: Cross-Lingual Transfer Learning (CLTL) offers a promising approach by leveraging knowledge from high-resource languages (source) to improve detection in low-resource languages (target).
The Gap: Prior surveys focused on general offensive language detection or specific monolingual scenarios. There was no comprehensive, systematic review dedicated exclusively to CLTL techniques in the context of offensive language detection, nor a unified taxonomy of the transfer strategies employed.

2. Methodology

The authors conducted a Systematic Literature Review (SLR) following the PRISMA framework to ensure replicability and transparency.

Search Protocol:
- Databases: ACL Anthology, Google Scholar, and DBLP.
- Keywords: 99 pairwise combinations of offensive language terms (e.g., "hate," "toxic," "cyberbully") and cross-lingual terms (e.g., "cross-lingual," "low resource," "transfer").
- Timeline: Search conducted up to July 29, 2023.
Selection Criteria:
- Inclusion: Studies describing cross-lingual approaches for general abusive content or subtypes (sexism, racism) involving knowledge transfer between languages.
- Exclusion: Studies focusing on cross-domain/platform transfer (non-linguistic), low-resource studies without transfer, competition reports, and dissertations.
Final Corpus: From an initial pool of 648 records, 67 relevant papers were selected for deep analysis.
Analysis Dimensions: The review categorized studies based on:
1. Multilingual datasets and resources used.
2. CLTL transfer levels (Instance, Feature, Parameter).
3. Specific transfer strategies (e.g., Zero-shot, Joint Learning).
4. Challenges and future directions.

3. Key Contributions

The paper makes four primary contributions to the field:

A. Comprehensive Dataset and Resource Analysis

Dataset Landscape: Analyzed 82 datasets used in the reviewed literature.
- Languages: Covers 32 languages across 10 families. English is dominant (25.95%), followed by Hindi (9.16%) and German (6.87%).
- Sources: Primarily Twitter (47.19%), YouTube, and Facebook.
- Labeling: 76% of datasets use binary labels (offensive/non-offensive), while others include fine-grained categories, attack targets, or intensity scores.
- Availability: 80% of datasets are publicly available.
Competitions: Summarized 24 major shared tasks (e.g., SemEval, HASOC, GermEval) that have driven progress in this domain.
Resources: Cataloged essential cross-lingual resources including multilingual lexicons (e.g., HurtLex), parallel corpora, machine translation tools, and pre-trained models.

B. Novel Taxonomy of CLTL Approaches

The authors propose a fine-grained taxonomy based on "what is transferred", categorizing 67 papers into three levels:

Instance Transfer: Transferring data elements (texts or labels).
- Label-based: Annotation projection, pseudo-labeling.
- Text-based: Machine translation (MT), text alignment, back-translation.
Feature Transfer: Transferring linguistic representations.
- Distributional: Multilingual Word Embeddings (e.g., MUSE, FastText, GloVe).
- Contextualized: Sentence embeddings from PLMs (e.g., LASER, LabSE).
- Retrofitting: Injecting domain knowledge (e.g., hate-specific dictionaries) into embeddings.
Parameter Transfer: Transferring model weights/parameters.
- Zero-Shot Transfer: Training on source only, testing on unseen target (e.g., Single Source $\to$ Single Target).
- Joint Learning: Training on source and target simultaneously (Early, Model-level, or Late fusion).
- Cascade Learning: Pre-training on source, then fine-tuning on target (including few-shot and meta-learning).

C. Identification of Challenges

The review synthesizes critical bottlenecks:

Language: Low-resource languages, typological distance (e.g., English $\to$ Chinese is harder than English $\to$ German), code-mixing, and cultural variations in defining "offense."
Dataset: Inconsistent definitions of hate speech, data scarcity, class imbalance, annotation bias, and dataset obsolescence due to evolving slang.
Approach: Limitations of Multilingual Pre-trained Language Models (mPLMs) in dialects/minority languages, poor interpretability of "black box" models, and translation errors diminishing toxicity perception.

D. Future Directions

The paper outlines actionable research paths:

Datasets: Need for culturally diverse, dialect-specific, and multi-platform datasets.
Annotation: Use of active learning and culturally contextualized guidelines to reduce bias.
Models: Moving from general mPLMs to domain-specific (e.g., XLM-T) and language-family-specific (e.g., MuRIL, AraBERT) models.
LLMs: Investigating the potential of Large Language Models (LLMs) for zero-shot/few-shot detection via prompting, while acknowledging their computational cost and lack of specialization compared to fine-tuned PLMs.

4. Results and Findings

Dominant Paradigm Shift: The field has shifted from Feature-based (using handcrafted embeddings) and Instance-based (relying heavily on translation) approaches to Parameter-based transfer.
Model Performance: Multilingual Pre-trained Language Models (mPLMs) like mBERT and XLM-R are the current state-of-the-art, accounting for the majority of recent studies. They are used in all three transfer scenarios (Zero-shot, Joint, Cascade).
Transfer Effectiveness:
- Zero-shot transfer is effective for typologically similar languages but degrades significantly for distant languages.
- Fine-tuning with even a small amount of target data (few-shot) consistently outperforms pure zero-shot approaches.
- Hybrid strategies (e.g., combining instance transfer for alignment with parameter transfer for fine-tuning) often yield the best results.
Language Bias: Current research is heavily skewed toward Indo-European languages. African, Indigenous, and smaller Asian languages remain underrepresented.
Reproducibility: Only ~38% of the reviewed papers made their source code publicly available, hindering reproducibility.

5. Significance

First Holistic Review: This is the first survey to exclusively focus on the cross-lingual aspect of offensive language detection, filling a critical gap in the literature.
Practical Guide: By providing a structured taxonomy and summarizing 82 datasets and 24 competitions, it serves as a foundational resource for researchers entering the field.
Ethical Awareness: The paper emphasizes the ethical implications of CLTL, including the risk of propagating biases across languages, the need for cultural sensitivity in annotation, and the dangers of automated censorship.
Roadmap for Future Work: It clearly delineates the path forward, advocating for the integration of cultural knowledge, the development of specialized PLMs for low-resource languages, and the responsible exploration of LLMs.

In conclusion, the paper establishes that while CLTL is a powerful tool for mitigating data scarcity in offensive language detection, significant challenges remain regarding linguistic diversity, cultural nuance, and model generalizability. The transition toward parameter-based transfer using specialized multilingual models represents the current frontier of the field.