Mangaluru: At a time when efforts are underway to secure recognition for Tulu as the second official language of the state, Prathamesh Devadiga, an eighth-semester BTech (CSE) student at PES university, Bengaluru, is working towards strengthening the language’s digital future, by working on a project that enables large language models (LLMs) to generate Tulu.Prathamesh told TOI that his work uses a structured five-layer prompt to get models such as GPT, Gemini, and Llama to produce grammatically correct Tulu, a low-resource dravidian language, without any fine-tuning.In his research paper, ‘Making Large Language Models Speak Tulu: Structured Prompting for an Extremely Low-Resource Language’, published in ‘Lossfunk Letters’, Prathamesh explores a practical approach to language preservation through the use of AI. “While LLMs are transforming communication worldwide, their benefits remain unevenly distributed, with digitally rich languages gaining most support and low-resource languages such as Tulu continuing to remain on the margins,” he said.Prathamesh said that though the language is spoken by nearly 2 million people, Tulu has a very limited digital presence. As a result, mainstream language models often fail to generate genuine Tulu even when explicitly instructed to do so, and tend to respond in Kannada. “This happens because Tulu and Kannada often share the same script and have overlapping vocabulary, and Kannada has a much stronger presence online,” he said.He said that a structured prompt was built around 2,800 tokens that enables models such as GPT and Gemini to generate Tulu instead of defaulting to Kannada. Prathamesh said vocabulary contamination was reduced from 80% to just 5%, while grammatical accuracy reached 85%. Crossmodel analysis showed that negative constraints consistently improved performance by 12 to 18% points, while grammar documentation contributed gains between 8 and 22% points depending on the model architecture.Calling it significant for language preservation in the AI era, Prathamesh said he is now in touch with developers to take the project further.
