Named Entity Recognition in Bengali and Hindi Using MuRIL and Conditional Random Fields

Named Entity Recognition (NER) is a sequence labelling task of Natural Language Processing (NLP) which aims to assign words to a pre-established list of named entity classes. Although a variety of NER models have been proposed to date, researchers have yet to find a viable solution for a poorly resourced and inflectional language like Bengali and Hindi. Additionally, many existing NER methods for low-resourced Indian language heavily rely on handcrafted features. This paper uses a multilingual neural language model called Multilingual Representations for Indian Languages (MuRIL) to extract deep features and Conditional Random Fields (CRF) for name entity tagging. We test the proposed model on two distinct Bengali NER datasets and a Hindi dataset. For the Bengali dataset-1, our proposed model achieves a Message Understanding Conference (MUC) F1 score of 75.02% which is approximately 3% higher than the MUC F1 score obtained by a recently published Bengali NER model. For the Bengali dataset-2, our proposed model achieves a MUC F1 score of 84.19%. For the Hindi dataset our proposed model achieves a MUC F1 score of 89.79% and it outperforms some existing deep learning-based Hindi NER models. We have also evaluated the proposed models using the tag-level F1 score. For both Bengali and Hindi datasets, the proposed model also outperforms the existing models in terms of the tag level F1 score.