Benchmarking explainable offensive language detection in Somali with human-annotated rationales.

Area of Science:

• Natural Language Processing
• Computational Linguistics
• Artificial Intelligence Ethics

Background:

• Offensive language detection is crucial for online safety but lacks interpretable justification methods.
• Existing datasets often lack rationales, and models offer limited transparency, hindering trustworthy content moderation.
• Somali, a low-resource language, faces unique challenges in developing effective and interpretable moderation systems.

Purpose of the Study:

• To introduce SomOffXplain, an interpretable framework for offensive language detection in Somali.
• To generate human-understandable explanations for offensive language classifications.
• To address the limitations of current models in interpretability and transparency for low-resource languages.

Main Methods:

• Developed SomOffXplain, a framework for span-level rationale extraction (word and phrase levels) in Somali.
• Constructed a new benchmark dataset of 10,175 Somali samples with human-annotated rationales.
• Evaluated SomOffXplain against fine-tuned pre-trained models (using LIME) and adapted large language models (LLMs) via few-shot and zero-shot prompting.

Main Results:

• SomOffXplain demonstrated superior explainability and predictive accuracy compared to baseline models.
• The proposed model exhibited higher plausibility and faithfulness in its generated rationales.
• Half of evaluated state-of-the-art LLMs failed to produce high-quality, human-aligned rationales; LIME-based methods were also weak explainers for Somali.

Conclusions:

• SomOffXplain enhances online safety and prevents harassment in under-resourced language communities.
• The framework improves the trustworthiness and transparency of AI systems for content moderation.
• This work provides a crucial step towards interpretable AI for diverse linguistic contexts.