Large Language Models for Devanagari Script Languages: Detecting Language, Hate Speech, and Targets in the CHiPSAL 2025 Shared Task

Core Concepts

This research paper presents a system for detecting language, hate speech, and targets of hate speech in Devanagari script languages, achieving competitive results in the CHiPSAL 2025 shared task by leveraging large language models like MuRIL, IndicBERT, and Gemma-2.

Abstract

• Bibliographic Information: Purbey, J., Pullakhandam, S., Mehreen, K., Arham, M., Sharma, D., Srivastava, A., & Kadiyala, R. M. (2024). 1-800-SHARED-TASKS @ NLU of Devanagari Script Languages: Detection of Language, Hate Speech, and Targets using LLMs. arXiv preprint arXiv:2411.06850v1.
• Research Objective: This paper describes the authors' system submitted to the CHiPSAL 2025 shared task, focusing on three sub-tasks: language identification in Devanagari script, hate speech detection, and hate speech target identification.
• Methodology: The researchers experimented with various large language models (LLMs) and ensembles, including MuRIL, IndicBERT V2, Gemma-2, Llama 3.1, and Mistral Nemo. They fine-tuned these models on a dataset of Devanagari script text and used techniques like focal loss to address class imbalance. For Sub-task A (language detection), both decoder-only and BERT-based models were used. Sub-task B (hate speech detection) also employed both types of models, with focal loss applied to BERT-based models to handle class imbalance. For Sub-task C (target detection), only decoder models were used due to the poor performance of BERT-based models in initial tests.
• Key Findings: The system achieved an F1 score of 0.9980 for language detection, 0.7652 for hate speech detection, and 0.6804 for target detection. The results demonstrate the effectiveness of LLMs, particularly Gemma-2 and IndicBERT V2, in handling complex NLP tasks for under-resourced languages.
• Main Conclusions: The study highlights the potential of LLMs in advancing language technologies for South Asian languages and emphasizes the importance of developing tailored approaches for specific NLP tasks.
• Significance: This research contributes to the growing field of hate speech detection and offers valuable insights for building inclusive language processing tools for diverse linguistic landscapes.
• Limitations and Future Research: The authors acknowledge the limitations posed by the relatively small datasets and computational constraints. Future research could explore larger, more diverse datasets and investigate more computationally efficient models.

Stats

The system achieved an F1 score of 0.9980 for language detection.
The F1 score for hate speech detection was 0.7652.
Target detection yielded an F1 score of 0.6804.
South Asian languages represent over 700 languages, 25 major scripts, and approximately 1.97 billion people.

Quotes

"Large language models (LLMs) have revolutionized natural language processing (NLP) yet South Asian languages remain largely underrepresented within these advancements..."
"Our work contributes to advancing language technology in South Asia, aiming for inclusivity and deeper understanding across diverse linguistic landscapes."

Key Insights Distilled From

1-800-SHARED-TASKS @ NLU of Devanagari Script Languages: Detection of Language, Hate Speech, and Targets using LLMs

by Jebish Purbe... at arxiv.org 11-12-2024

https://arxiv.org/pdf/2411.06850.pdf

Deeper Inquiries

How can the fairness and bias of hate speech detection models be evaluated and improved, especially considering the nuances of different languages and cultural contexts?

Evaluating and mitigating bias in hate speech detection models, especially across diverse languages and cultures, is a multifaceted challenge requiring a multi-pronged approach:
Evaluation:

Dataset Analysis:

Representation Analysis: Scrutinize the dataset for representation biases. Are certain demographic groups over-represented or under-represented in the data, particularly within hate speech and non-hate speech categories?
Label Bias Analysis: Examine the labels assigned to the data. Do the annotations themselves reflect biases? This requires careful consideration of cultural context, as what constitutes hate speech can vary significantly.

Model Evaluation Metrics:

Go Beyond Accuracy: Don't solely rely on overall accuracy. Utilize metrics like:

Precision and Recall: Measure how well the model identifies true positives (hate speech) while minimizing false positives (misclassifying non-hate speech).
F1-Score: Provides a balanced measure of precision and recall.
Disaggregated Evaluation: Analyze performance across different demographic groups and language variations. Are there significant differences in error rates, indicating potential bias?

Counterfactual Analysis:

Perturb Sensitive Attributes: Systematically alter potentially sensitive attributes (e.g., change names to represent different ethnicities) in the input text and observe how the model's predictions change. This helps identify if the model is overly reliant on these attributes.

Improvement:

Data Collection and Annotation:

Diverse Data Sources: Gather data from a wide range of sources to ensure representation across demographics, dialects, and language styles.
Bias-Aware Annotation Guidelines: Develop detailed guidelines for annotators that explicitly address potential biases and provide clear examples of hate speech within specific cultural and linguistic contexts.
Multiple Annotators: Employ multiple annotators per data point and resolve disagreements through discussion or adjudication to minimize individual bias.

Model Development:

Bias Mitigation Techniques: Explore techniques like:

Adversarial Training: Train the model to be less sensitive to sensitive attributes while maintaining accuracy.
Debiasing Word Embeddings: Identify and mitigate biases present in the word embeddings used by the model.

Explainability and Interpretability: Utilize models or techniques that offer insights into why a particular prediction was made. This transparency helps identify and address potential biases in the decision-making process.

Continuous Monitoring and Evaluation:

Real-World Performance: Continuously monitor the model's performance after deployment, paying close attention to potential disparities in how it affects different user groups.
Feedback Mechanisms: Establish channels for user feedback to gather insights into potential biases and areas for improvement.

Cultural and Linguistic Nuances:

Language-Specific Expertise: Involve linguists and cultural experts familiar with the target languages and dialects to ensure accurate understanding and interpretation of potentially offensive language.
Contextual Understanding: Develop models that consider the broader context of the conversation, including cultural references, humor, and sarcasm, to avoid misinterpreting intent.
Addressing fairness and bias is an ongoing process. Regular evaluation, model updates, and collaboration with communities impacted by these technologies are crucial for responsible and ethical hate speech detection.

Could the integration of other modalities, such as audio or visual information, alongside text, enhance the accuracy and robustness of hate speech detection systems?

Yes, integrating other modalities like audio and visual information with text can significantly enhance the accuracy and robustness of hate speech detection systems. Here's how:
1. Capturing Nuances Beyond Text:

Sarcasm and Tone: Text alone often fails to convey sarcasm, humor, or intent, leading to misinterpretations. Audio analysis can detect vocal tones, inflections, and emphasis, providing crucial context.
Visual Cues: Visuals accompanying text, such as images, memes, or videos, often amplify or contradict the message's sentiment. Image recognition can identify hate symbols, while video analysis can interpret facial expressions and gestures, adding valuable information.
2. Addressing Multimodal Hate Speech:

Beyond Text: Hate speech increasingly incorporates images, videos, and audio. Multimodal systems can detect hate speech expressed through visual metaphors, hateful symbols in images, or aggressive vocal tones in audio messages.
3. Improved Accuracy and Robustness:

Cross-Modal Confirmation: When modalities agree, it strengthens the confidence in hate speech detection. For example, if text expresses hate and the audio conveys anger, it's more likely to be flagged correctly.
Handling Ambiguity: When one modality is ambiguous, others can provide clarity. For instance, a sarcastic text message paired with a laughing voice is less likely to be misclassified as hateful.
Robustness to Textual Manipulation: Users sometimes try to evade text-based filters by misspelling words or using special characters. Multimodal systems are more robust to such manipulations as they analyze additional information.
Examples of Multimodal Integration:

Social Media Posts: Analyzing text, images, and videos in a tweet or post.
Live Streams: Detecting hate speech in real-time by combining audio, video, and chat messages.
Online Games: Monitoring in-game chat alongside player voice communication.
Challenges:

Data Complexity: Collecting and annotating multimodal data is more challenging and expensive than text-only data.
Computational Resources: Processing and analyzing multimodal data requires significant computational power.
Privacy Concerns: Using audio and video data raises privacy concerns, requiring careful consideration of data security and user consent.
Despite the challenges, multimodal hate speech detection holds immense potential. As technology advances and datasets grow, we can expect more sophisticated systems that leverage the power of multiple modalities to create safer and more inclusive online spaces.

What are the potential societal implications of widespread adoption of hate speech detection technologies, and how can we ensure their ethical and responsible use in online platforms?

The widespread adoption of hate speech detection technologies presents a complex interplay of potential benefits and risks. While these technologies can contribute to a safer and more inclusive online environment, their ethical and responsible use requires careful consideration and ongoing dialogue.
Potential Positive Implications:

Reduced Toxicity: By identifying and mitigating hate speech, these technologies can help create a more respectful and inclusive online environment, fostering healthier discussions and reducing the harmful effects of online harassment.
Protection of Vulnerable Groups: Automated detection can help protect marginalized communities disproportionately targeted by online hate speech, offering them a degree of protection and recourse.
Increased Accountability: The presence of these systems can deter individuals from engaging in hate speech, promoting greater accountability for online behavior.
Potential Negative Implications:

Censorship and Free Speech: Overly aggressive filtering can stifle legitimate expression and dissent, particularly if definitions of hate speech are too broad or culturally insensitive.
Bias and Discrimination: As discussed earlier, biased datasets or models can lead to the unfair targeting of specific groups or the suppression of their voices.
Erosion of Trust: Lack of transparency in how these systems work and potential for errors can erode trust in online platforms and create a chilling effect on free expression.
Ensuring Ethical and Responsible Use:

Transparency and Explainability: Platforms should be transparent about their hate speech policies, the technologies they employ, and how these systems make decisions. Explainable AI can help users understand why content is flagged, increasing fairness and accountability.
Human Oversight and Appeal Mechanisms: Automated systems should not be the sole arbiters of hate speech. Human review is crucial for handling complex cases, cultural nuances, and providing users with avenues for appeal if their content is incorrectly flagged.
Contextual Awareness and Nuance: Models should consider the broader context of conversations, including intent, relationships between users, and cultural factors, to avoid misinterpretations and censorship of legitimate expression.
Ongoing Evaluation and Improvement: Continuous monitoring, evaluation, and improvement of these systems are essential to address biases, adapt to evolving language, and ensure fairness.
Public Discourse and Regulation: Open discussions involving policymakers, technology companies, researchers, and the public are crucial for establishing clear guidelines, ethical frameworks, and potential regulations for the development and deployment of hate speech detection technologies.

Balancing Act:
The challenge lies in striking a balance between protecting users from harm and upholding freedom of expression. Ethical and responsible use requires a multi-stakeholder approach, prioritizing transparency, fairness, and ongoing dialogue to harness the benefits of these technologies while mitigating potential risks.