Transformer-Based Named Entity Recognition Model—Tamil Language

This work presents different transformer-based approaches to the named entity recognition problem in the Tamil language. The WikiANN-ta dataset is used for training, validating and testing the model. Person, organization and location tags are annotated in the dataset. Context features of Tamil languages are extracted by the BERT pre-trained models. In this task, five significant BERT variations are explored, (1) mBERT (multilingual BERT), (2) XLM-RoBERTa, (3) Indic BERT, (4) MuRIL base and (5) + MuRIL large. Context features extracted by mBERT/XLM-R/Indic BERT/ MuRIL base/MuRIL large are sent to the feedforward layer with SoftMax to identify named entities. Two models out of these five BERT-based models achieve state-of-the-art results in the test dataset of WikiANN-ta (1) NER model with MuRIL base, (2) NER model with MuRIL-large. MuRIL base is proposed for better run time, and MuRIL-large is proposed for better accuracy.