Evaluating Ensemble-Based Classifiers for Khasi Named Entity Recognition with Data Imbalance Constraints
Science & Technology Journal · 27.05.2026 · 14 (Online First)
DOI: 10.22232/stj.2026.2 ↗Kurz gefasst
Fachartikel in der Zeitschrift Science & Technology Journal, online erschienen am 27.05.2026. Verfasst von Amitabha Nath und 2 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Named Entity Recognition (NER), a core task of NLP, has been the subject of much research in resource-rich languages. However, little has been attempted with respect to low-resource languages such as Khasi. Based on our previous efforts, the existing works on Khasi NER applied transformer models such as RoBERTa, XLM-RoBERTa, and BERT-base-multilingual-cased using transfer learning (TL) and achieved an F1 score of 91.25% with RoBERTa. In this paper, ensemble methods are investigated to improve performance even further. We used a novel Khasi named entity recognition dataset annotated in span-based and converted to IOB2 formats for token-level predictions for both TL and ensemble methods. Ensemble methods, including majority voting, weighted voting, and stacking with logistic regression and XGBoost, were employed. Two settings were tested, one where two models are implemented (RoBERTa and XLM-RoBERTa) and the other with three models, with the addition of BERT-base-multilingual-cased, where these models act as learners. The ensemble of three models with XGBoost scored the highest F1 score of 94.06%, which is an improvement of +2.81% over the previous best. It also achieved significant improvements on rare entities, such as WORK_OF_ ART, LANGUAGE, ORDINAL, and EVENT, although the dataset was skewed toward the PERSON class. The results show how ensemble methods can improve NER performance on imbalanced datasets and low-resource languages such as Khasi.
Zitieren
APA
Nath, A., Hoojon, R., & Lyngdoh, S. (2026). Evaluating Ensemble-Based Classifiers for Khasi Named Entity Recognition with Data Imbalance Constraints. Science & Technology Journal, 14(Online First). https://doi.org/10.22232/stj.2026.2
Vancouver
Nath A, Hoojon R, Lyngdoh S. Evaluating Ensemble-Based Classifiers for Khasi Named Entity Recognition with Data Imbalance Constraints. Science & Technology Journal. 2026;14(Online First). doi:10.22232/stj.2026.2
BibTeX
@article{Nath2026,
author = {Nath, Amitabha and Hoojon, Ransly and Lyngdoh, Saralin},
title = {Evaluating Ensemble-Based Classifiers for Khasi Named Entity Recognition with Data Imbalance Constraints},
journal = {Science & Technology Journal},
year = {2026},
volume = {14},
number = {Online First},
doi = {10.22232/stj.2026.2}
}
EUVIMED