L3CubeMahaNER: A Marathi Named Entity Recognition Dataset & BERT Models

Dataset Overview

L3CubeMahaNER is a publicly released corpus for Named Entity Recognition (NER) in Marathi, one of the major Indic languages spoken by over 80million people. The resource was curated by the Language Technologies Research Group at the L3Cube research lab and released under a CCBY4.0 license, allowing unrestricted academic and commercial use.

The primary goal of the dataset is to fill the longstanding gap of highquality, domaindiverse annotated text for Marathi. While several Indianlanguage NER corpora exist for Hindi, Bengali, or Tamil, Marathi has lagged behind, limiting progress on downstream applications such as information extraction, question answering, and conversational agents.

The corpus consists of 10,000 sentences selected from news articles, government reports, Wikipedia, and social media, providing a realistic mix of formal and informal language.

Annotation Scheme

The annotation follows the widely adopted BIO tagging format (Beginning, Inside, Outside) and covers the four coarsegrained entity types:

• PER Person names
• LOC Geographical locations (cities, states, landmarks)
• ORG Organizations, institutions, companies
• MISC Miscellaneous entities such as dates, events, titles, and product names

Example (transliterated):

        LOCB          LOCI           O        O        PERB       PERI    PERI.

Annotators were native Marathi speakers with a background in linguistics. Each sentence was doubleannotated and disagreements were resolved through a third adjudicator, ensuring a Cohens Kappa of 0.91 (excellent agreement).

Key Statistics

The split follows the conventional 80/10/10 ratio:

• Training set 8,000 sentences
• Validation set 1,000 sentences
• Test set 1,000 sentences

BERT for Marathi NER

Transformerbased language models have become the defacto standard for sequence labeling tasks. For Marathi, several BERTstyle models are available:

1. IndicBERT A multilingual model pretrained on 12 Indic languages including Marathi.
2. MuRIL (Multilingual Representations for Indian Languages) Developed by Google, trained on a massive Indianlanguage corpus.
3. MarathiBERT A monolingual BERT trained exclusively on Marathi Wikipedia, news, and web text ( 30M sentences).

All three were finetuned on the L3CubeMahaNER training split using a tokenclassification head. The typical hyperparameters were:

• Learning rate: 3e5
• Batch size: 16
• Epochs: 4
• Optimizer: AdamW with weight decay 0.01

Why BERT Works Well for Marathi

• Subword tokenisation WordPiece handles the agglutinative morphology of Marathi, splitting rare forms into known morphemes.
• Contextual embeddings The same word can receive different representations based on surrounding words, crucial for disambiguating homographs.
• Transfer learning Pretraining on large unlabelled Marathi text infuses syntactic and semantic knowledge that is transferred to the NER task.

Experimental Results

The following table summarises the microaveraged precision, recall and F1score achieved on the official test set.

Adding a Conditional Random Field (CRF) on top of the BERT token classifier yields the best performance, confirming that modeling label dependencies (e.g., ensuring ILOC follows BLOC) improves consistency.

Error analysis revealed two dominant failure modes:

• Ambiguous nouns Words that can be both common nouns and proper names (e.g., meaning sun vs. a persons surname).
• Codeswitching Frequent English insertions in informal text cause the model to miss entities or mislabel them.

How to Use the Dataset

The dataset can be downloaded from the official GitHub repository:

git clone https://github.com/L3Cube-MahaNER/mahaner.git

Inside the repository you will find three JSONlines files (train.jsonl, dev.jsonl, test.jsonl) where each line contains a JSON object:

{    "sentence": "    ",    "labels": ["LOCB", "LOCI", "O", "O", "O", "O"]}        

A minimal PyTorch training script using Transformers is provided in the examples/ folder. The script demonstrates:

1. Loading the dataset with datasets.load_dataset
2. Tokenising with the appropriate tokenizer (e.g., MarathiBertTokenizer)
3. Aligning BIO labels to subword tokens
4. Finetuning the model and evaluating via seqeval

Researchers can also experiment with other architectures (BiLSTMCRF, spaCy, Flair) by converting the JSONL to CoNLL2003 format using the helper script jsonl_to_conll.py.

Community Contributions

Since its release, the community has extended L3CubeMahaNER with:

• Domainspecific subsets (legal, medical)
• Finegrained entity types (e.g., DATE, TIME)
• Crosslingual evaluation using MarathiHindi parallel corpora

Contributions are welcomed via pull requests and issue tracking on the GitHub repo.