L3Cube-MahaPOS: A Marathi Part-of-Speech Tagging Dataset and BERT Models

Part-of-Speech (POS) tagging is a fundamental NLP task that provides the structural metadata necessary for downstream applications like machine translation and syntactic parsing. While high-resource languages benefit from mature tools and large-scale benchmarks, Marathi—a language spoken by over 83 million people—has historically lacked standardized, high-quality annotated corpora. This paper addresses this gap by introducing L3Cube-MahaPOS, a gold-standard dataset designed to support robust POS tagging research for Marathi.

The L3Cube-MahaPOS dataset consists of 32,354 manually annotated sentences sourced from diverse Marathi news portals. The authors implemented a rigorous four-stage preprocessing pipeline, including Unicode normalization, Devanagari-aware tokenization, and noise filtering, to ensure data consistency. Annotation followed a 16-tag scheme aligned with the Universal Dependencies (UD) framework, with native-speaker annotators resolving ambiguities through iterative guideline updates and majority-label adjudication.

The authors evaluated six model families to establish a reproducible baseline for the field. These ranged from classical statistical models (HMM, CRF) to deep learning architectures (BiLSTM, BiLSTM+CharCNN) and modern transformer-based models (MuRIL, MahaBERT-v2). The results show that transformer-based models significantly outperform traditional approaches, with the Marathi-specific MahaBERT-v2 achieving 88.67% token-level accuracy and a macro-F1 score of 81.67%. This study provides the community with a standardized benchmark and open-source model checkpoints to facilitate future development in Marathi NLP.