A Comparative Study: Machine Learning and Deep Learning based Approaches for Recognizing Named Entity from Bengali Corpora

View at publisher Summarise with VIRAT
Authors: Diza Sultana, Ankita Chakraborty, Anupam Mondal
Cited by: 0
References: 22

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

The development of NER systems in Bengali faces challenges owing to complex morphology, word order flexibility, and a dearth of well-annotated data sources. This work presents an empirical assessment of three Bengali NER approaches, namely Conditional Random Field (CRF), a Bidirectional Long Short-Term Memory (BiLSTM) network, and the multilingual transformer XLM-RoBERTa. Controlled experiments are conducted using a main Bengali NER dataset as well as adding a side dataset with an increased number of samples. The model performance is assessed through macro and micro averaged F1 scores. The CRF model exhibits consistent performance, achieving the best macro f1 as 0.8238 (micro f1 0.8735) in mainonly setting and the best macro f1 as 0.8116 (micro f1 0.8920) in combined setting in 10-fold cross validation setup. The BiLSTM model shows weaker performance due to zero prior pretrained contextual representations, achieving best macro$\text{f1}$as 0.36 and micro f1 as 0.3793 in main-only setting; and the best macro f1 as 0.38 and micro f 1 as 0.4036 in combined setting. The XLMRoBERTa yields the highest overall performance, with the best f1 scores of 0.8120 in validation set and 0.8090 in test set on the main dataset. Addition of side dataset gives similar results often negatively affecting them. In our work, we find that transformers are well-suited for Bengali NER annotation, while CRF models remain dependable in low-resource conditions.

Fields of Study & Subjects

Topic Modeling Natural Language Processing Techniques Handwritten Text Recognition Techniques

Publication details

Journal / source
International Conference on Electronics, Materials Engineering & Nano-Technology (IEMENTech)
Pages
1–6
Publication date
6 February 2026
Language
EN