Low-resource Machine Translation

Building corpora and MT systems for Nepali, Tamang, and other under-served Indic languages, from data pipelines to fine-tuned NLLB models and a live translation platform, plus data-scaling studies across Nepali, Hindi, and Tamang to English.

Funded by: Google Academic Research Award 2024

Code-Mixing

Studying Nepali–English code-mixed text: corpus construction from real-world speech transcripts, translation into pure Nepali, analysis of how code-mixing varies across profession groups, and benchmarking models on code-mixed data.

Funded by: UGC Nepal Faculty Research Award

Multilingual LLMs and AI Safety

Adapting open LLMs such as Gemma to Nepali through continued pre-training and post-training (SFT), and examining the reliability, robustness, safety, and guardrails of multilingual models for low-resource languages.

Legal Argument Mining

Identifying and classifying arguments in legal text, using parameter-efficient fine-tuning of LLMs with retrieval augmentation on the ECHR corpus, and domain-adapting small language models to Nepali law.

Commitment Mining

Detecting and classifying election candidates' commitments in Nepali political text: building the first gold-standard dataset of 2.4K+ commitments, benchmarking ML and DL models, and building Janchetana, a public app for tracking them.

Funded by: KU Research, Development, and Innovation Grant

Publications →  ·  Datasets, models & tools →