SampurNER: Fine-grained Named Entity Recognition dataset for 22 Indian Languages

We introduce SampurNER, a fine-grained named entity recognition (FgNER) dataset encompassing all 22 scheduled Indian languages spoken by more than two billion people across various countries.

We have proposed an entity-anchored machine translation (EaMaTa) framework that leverages the largest manually annotated English FgNER dataset, FewNERD, to create a large-scale FgNER dataset in 22 languages.

On average, the dataset comprises over 153k sentences, 354k entities, and 3.3M tokens in each language.

The languages covered are: Assamese (as), Bengali (bn), Bodo (brx), Dogri (doi), Gujarati (gu), Hindi (hi), Kannada (kn), Kashmiri (ks), Konkani (gom), Maithili (mai), Malayalam (ml), Manipuri (mni), Marathi (mr), Nepali (ne), Odia (or), Punjabi (pa), Sanskrit (sa), Santali (sat), Sindhi (sd), Tamil (ta), Telugu (te), and Urdu (ur).

Various rigorous analyses and human evaluations confirm the high quality of the dataset and demonstrate the effectiveness of the entity-anchored machine translation (EaMaTa) framework with up to 9% increase in F1-score against the current state-of-the-art.

Read the extended version

EaMaTa Framework overview

Figure: Overview of the EaMaTa Framework.

SampurNER Dataset Statictics

Language	Train set			Development set			Test set
	Sentences	Entities	Tokens	Sentences	Entities	Tokens	Sentences	Entities	Tokens
Assamese (as)	107,249	237,260	2,194,925	15,438	34,560	318,105	30,658	67,466	625,870
Bengali (bn)	119,296	287,264	2,484,304	17,513	42,877	368,063	33,374	79,340	689,690
Bodo (brx)	117,659	262,792	2,354,696	16,762	37,496	336,269	33,615	74,576	672,246
Dogri (doi)	112,329	264,154	2,885,149	17,619	42,526	459,537	34,931	82,597	903,796
Gujarati (gu)	126,581	315,919	2,828,298	18,122	45,431	406,929	28,959	69,207	619,889
Hindi (hi)	124,887	290,192	3,298,116	17,882	41,824	457,573	35,713	82,440	908,513
Kannada (kn)	115,565	266,523	2,083,241	16,962	39,781	308,326	26,327	59,365	453,817
Kashmiri (ks)	123,679	288,544	2,910,937	17,417	40,350	408,053	35,106	81,181	823,040
Konkani (gom)	83,415	182,806	1,637,018	12,276	27,262	243,817	23,759	51,483	463,980
Maithili (mai)	108,826	256,701	2,763,005	10,224	22,706	245,657	19,899	43,530	472,498
Malayalam (ml)	91,743	199,485	1,504,839	15,608	35,140	265,049	23,480	50,319	377,213
Manipuri (mni)	110,068	246,084	2,264,925	15,561	34,869	321,556	31,463	69,739	644,709
Marathi (mr)	125,543	309,220	2,614,024	17,650	43,407	367,882	36,237	89,295	754,851
Nepali (ne)	125,695	311,439	2,661,064	18,252	45,778	389,382	35,498	87,112	747,802
Odia (or)	118,633	289,943	2,427,051	18,090	45,247	376,152	32,477	78,893	657,395
Punjabi (pa)	96,986	234,436	2,348,393	17,655	44,415	443,788	36,920	92,655	928,798
Sanskrit (sa)	69,581	152,269	1,214,021	10,043	22,175	176,574	19,729	42,643	341,208
Santali (sat)	87,650	153,533	2,223,951	12,526	22,159	312,706	24,921	43,264	619,556
Sindhi (sd)	90,362	214,371	2,218,078	17,221	42,845	440,340	32,159	78,317	809,085
Tamil (ta)	96,004	216,285	1,711,203	10,702	23,542	183,893	25,160	55,927	441,141
Telugu (te)	85,893	193,425	1,505,321	16,790	39,909	309,345	21,729	47,988	372,946
Urdu (ur)	122,794	298,069	298,069	17,570	43,205	465,417	35,198	35,198	929,427

Citation

If you use this dataset, please cite the following paper:

@inproceedings{kaushik2026sampurner,
  title={SampurNER: Fine-grained Named Entity Recognition dataset for 22 Indian Languages},
  author={Kaushik, Prachuryya and Anand, Ashish},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={40},
  year={2026}
}