NerAxom-MGF: Leveraging Multi-Granularity Fusion for Assamese Named Entity Recognition
Information Retrieval Research · 2026
DOI: 10.54195/irrj.26376 ↗Kurz gefasst
Fachartikel in der Zeitschrift Information Retrieval Research, 2026. Verfasst von Punam Sarmah und 3 weiteren Autor:innen. Die Arbeit nennt 51 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
We introduce MGF-NER, a multi-granularity fusion architecture that integrates character-level, subword-level, and entity-span-level representations for the same task using NerAxom dataset. NerAxom comprises 102,857 tokens across 4,178 sentences annotated under the BIO scheme with seven entity categories: Person, Location, Organization, Date, Number, Occupation, and Work-of-Art. MGF-NER combines a Character-level Convolutional Neural Network that captures morpheme-level cues attenuated by subword fragmentation, a pre-trained MuRIL transformer for cross-lingually pretrained contextual semantics, an Entity Span Attention module for multi-word boundary detection, and a Conditional Random Field decoder that enforces valid BIO transitions. Under a seeded five-run evaluation protocol on the NerAxom test set, MGF-NER achieves a mean micro-averaged F1 of 85.45 ± 0.46, outperforming the strongest single-encoder baseline MuRIL (83.88±0.79) by 1.57 F1 points; the improvement is statistically significant under a paired sentence-level Wilcoxon signed-rank test (p = 0.005). We benchmark MGF-NER against nine baselines spanning traditional feature-based sequence labelling (CRF), a recurrent baseline (BiLSTM-CRF), language-agnostic multilingual transformers (mBERT, XLM-RoBERTa base and large, DistilBERT-multi), a script-similar Bangla-specialised transformer (Bangla-BERT), and Indic-specialised transformers (IndicBERTv2). Single-component and compound-component ablations, together with per-category ablation, decompose the gain across the four architectural components, and error analysis characterises the residual error profile.
Zitieren
APA
Sarmah, P., Lahkar, M. P., Kalita, S., & Sharma, U. (2026). NerAxom-MGF: Leveraging Multi-Granularity Fusion for Assamese Named Entity Recognition. Information Retrieval Research. https://doi.org/10.54195/irrj.26376
Vancouver
Sarmah P, Lahkar MP, Kalita S, Sharma U. NerAxom-MGF: Leveraging Multi-Granularity Fusion for Assamese Named Entity Recognition. Information Retrieval Research. 2026. doi:10.54195/irrj.26376
BibTeX
@article{Sarmah2026,
author = {Sarmah, Punam and Lahkar, Manash P and Kalita, Shobhanjana and Sharma, Utpal},
title = {NerAxom-MGF: Leveraging Multi-Granularity Fusion for Assamese Named Entity Recognition},
journal = {Information Retrieval Research},
year = {2026},
doi = {10.54195/irrj.26376}
}
EUVIMED