Socio-Cultural Alignment of Large Language Models for Sinhala

Principal Investigator: Nisansa de Silva

The proposed research is expected to break new ground by addressing the dual challenges of linguistic complexity and cultural awareness in Sinhala, a low-resource language largely overlooked in LLM research. While prior studies have focused on high-resource languages, this work pioneers the systematic creation of Sinhala-specific datasets that embed both linguistic and cultural nuances, an area where current benchmarks and datasets are nearly absent.
The proposal offers several innovative approaches. First, it introduces a taxonomy-driven method for developing culture-aware datasets that go beyond simple MCQ formats, incorporating sentiment analysis, translation, and safety tasks. Second, it proposes Sinhala-focused benchmarks for LLM safety, addressing culture-specific toxicity and hate speech, an underexplored dimension in current safety research. Finally, novel training strategies for cross-lingual and cross-cultural alignment will be explored using open-source models such as Llama, offering replicable and scalable contributions.
The likelihood of this research being published and cited internationally is high, given the global interest in low-resource languages, cultural competence, and AI safety. The work fills a critical gap and is aligned with the research priorities of top-tier NLP and AI conferences.
The potential benefits are wide-ranging. For Sinhala speakers, culturally competent LLMs can improve applications in education, healthcare, governance, and digital inclusion, ensuring technology reflects local realities. More broadly, the methods and frameworks developed can serve as a blueprint for other low-resource languages, advancing global efforts toward equitable and culturally aware AI systems.

Objectives:

• Analyse Sinhala-related benchmarks, identify strengths and weaknesses, and propose a taxonomy for building culture-specific common-sense datasets.
• Create Sinhala-focused multi-cultural datasets (~2000 samples), including sentiment analysis, translation, and safety tasks (~500 samples), reflecting cultural nuances.
• Develop methods to enhance Sinhala LLMs’ cross-cultural awareness, improve safety, and mitigate culture-related risks of using open-source models such as Llama.