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

Researchers have released L3Cube-MahaPOS, a gold-standard part-of-speech tagging dataset for Marathi containing 32,354 manually annotated sentences, alongside benchmarks for six model families [1][2]. Marathi is spoken by over 83 million people and ranks among the top twenty most spoken languages worldwide, yet it remains severely under-resourced in annotated corpora and standardised evaluation benchmarks [1][2]. The language presents distinct computational challenges, including rich morphology, relatively free word order, a lack of capitalisation conventions, and pervasive code-mixing with Hindi and English [1][2]. The L3Cube-MahaPOS corpus was annotated entirely by a team of Marathi-proficient annotators following a 16-tag Universal Dependencies-aligned scheme [1][2]. A structured preprocessing pipeline covering Unicode normalisation, Devanagari-aware tokenisation, and noise filtering was applied to ensure label consistency across all data splits [1][2]. The sentences were drawn from news text [1][2]. The researchers benchmarked the dataset across six model families: HMM, CRF, BiLSTM, BiLSTM+CharCNN, MuRIL, and the Marathi-specific transformer MahaBERT-v2 [1][2]. The best-performing system achieved 88.67% token-level accuracy and a macro-F1 of 81.67% over 15 evaluated tag classes [1][2]. Part-of-speech tagging is a foundational natural language processing task that underpins machine translation, information extraction, and syntactic parsing [1][2]. The release of the dataset, annotation guidelines, and trained model checkpoints is intended to foster further research in Marathi NLP [1][2].