TL;DR: Comprehensive benchmark to evaluate LLMs on low-resource Indic languages. Dataset publicly available on HuggingFace. Addresses critical gap in multilingual NLP evaluation.

TL;DR: Domain-aware multilingual lexicon generation for 6 Indian languages across 8 domains using routing-based architecture. Released benchmark with 75K+ translation pairs. Accepted at ACL Main Conference 2025.

TL;DR: ARISE iteratively induces rules and generates synthetic data for text classification via bootstrapping. Outperforms complex methods like contrastive learning across diverse domains and languages. Published at NAACL 2025 Findings.

TL;DR: DictDis disambiguates between multiple dictionary candidate translations in lexically constrained NMT. Achieves 2-3 BLEU point improvements across regulatory, finance, engineering, and health domains. Published at EMNLP 2024 Findings.

TL;DR: First comprehensive benchmark and dataset for English-Sanskrit machine translation. Addresses critical gap in classical language NLP. Published at LREC-COLING 2024.

TL;DR: A production-ready MT post-editing tool used by 100+ translators to translate technical content into Indian languages. Won Best Paper Award at CODS-COMAD 2023. Impact: First batch of engineering books translated using UDAAN were released by the President of India.

TL;DR: AMAL learns instance-specific weights using meta-learning to optimally combine auxiliary losses in supervised learning. Provides significant gains in knowledge distillation and rule denoising. Published at AAAI 2023.

TL;DR: SPEAR combines semi-supervised learning with data programming to improve noisy labeling functions. Significantly outperforms state-of-the-art on seven datasets by jointly learning from rules and labeled data. Published at ACL 2021 Findings.

TL;DR: Jointly learns classifier parameters and hyperbolic label embeddings for hierarchical multi-label classification without assuming known label hierarchy. Achieves state-of-the-art results by capturing hierarchical structure in hyperbolic space. Published at EACL 2021.