xlm_latin-Multilingual_detection_of_hate_speech

GhadeerALbadani/xlm_latin-Multilingual_detection_of_hate_speech: Multilingual Hate Speech Detection Using Unified Latin Script

Model Details

Model Description

XLM_Latin is a multilingual Transformer-based model fine-tuned for multilingual hate speech detection using text standardized into a unified Latin script representation.

The model was fine-tuned using eight languages:

• Arabic
• Hebrew
• Persian
• Russian
• Spanish
• Bengali
• Chinese
• Korean

Before fine-tuning, the text from these languages was converted from their original writing systems into a unified Latin script through transliteration.

The model was then trained jointly on the eight standardized languages and evaluated individually on each language.

The classification task is binary:

• Hate Speech

• Non-Hate Speech

• Developed by: Ghadeer Albadani

• Model type: Multilingual Transformer-based sequence classification model

• Task: Multilingual Hate Speech Detection

• Languages: Arabic, Hebrew, Persian, Russian, Spanish, Bengali, Chinese, Korean

• Writing System: Unified Latin Script

• Method: Transliteration-based multilingual learning

• Classification: Binary text classification

• Framework: PyTorch / Hugging Face Transformers

• Fine-tuned from: XLM-RoBERTa

• License: Please refer to the license of the underlying pretrained model and datasets

Model Sources

• Repository: https://huggingface.co/GhadeerALbadani/XLM_Latin
• Task: Multilingual Hate Speech Detection
• Approach: Transliteration-based multilingual learning
• Training Scenario: Multilingual fine-tuning using eight languages standardized to Latin script

Uses

Direct Use

XLM_Latin is designed for binary hate speech classification on multilingual text that has been converted into the same unified Latin-script representation used during model training.

The model can be used for:

• Multilingual hate speech detection
• Cross-lingual NLP research
• Transliteration-based text classification
• Multilingual content analysis
• Research on unified writing systems
• Hate speech detection across different writing systems

Downstream Use

The model can be further adapted for:

• Social media hate speech detection
• Multilingual content moderation research
• Cross-lingual text classification
• Low-resource language NLP
• Offensive language detection
• Multilingual NLP applications

For downstream applications, the same preprocessing and transliteration procedure used during training should be applied to the input data.

Out-of-Scope Use

The model should not be used as the sole decision-making mechanism for:

• Legal or criminal decisions
• Automated punishment of individuals
• Permanent social media account suspension
• Political or social profiling
• High-stakes content moderation without human review

The model is primarily intended for research and automated classification assistance.

Bias, Risks, and Limitations

Hate speech detection is a context-sensitive task influenced by language, culture, dialect, topic, and social context.

Although transliteration provides a unified writing representation, it does not eliminate linguistic differences between languages.

The evaluation results demonstrate that model performance varies across languages.

The model achieved the following performance:

• Bengali: Macro F1 = 0.85
• Spanish: Macro F1 = 0.81
• Persian: Macro F1 = 0.80
• Russian: Macro F1 = 0.77
• Chinese: Macro F1 = 0.75
• Arabic: Macro F1 = 0.72
• Hebrew: Macro F1 = 0.72
• Korean: Macro F1 = 0.62

Potential limitations include:

• Transliteration ambiguity.
• Language-specific lexical and semantic differences.
• Difficulty with sarcasm and implicit hate speech.
• Difficulty with slang and informal social media language.
• False positives and false negatives.
• Domain shift between training and real-world data.
• Reduced performance on languages not represented during fine-tuning.

Recommendations

Users should evaluate the model separately for each target language and application domain.

For real-world applications, predictions should be combined with:

• Human review
• Contextual analysis
• Language-specific validation
• Periodic model evaluation
• Bias and fairness assessment

The model should be considered a decision-support system rather than a fully autonomous moderation system.

How to Get Started with the Model

The model can be loaded using the Hugging Face Transformers library.

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

model_name = "GhadeerALbadani/xlm_latin-Multilingual_detection_of_hate_speech"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

text = "your transliterated text here"

inputs = tokenizer(
    text,
    return_tensors="pt",
    truncation=True,
    padding=True
)

with torch.no_grad():
    outputs = model(**inputs)

prediction = torch.argmax(outputs.logits, dim=-1).item()

if prediction == 1:
    print("Hate Speech")
else:
    print("Non-Hate Speech")