text
stringlengths
0
255
जाँदा O
जाँदै O
, O
बाह्रबुंदे O
अन्तर्गत O
को O
मनपरी O
तन्त्र O
रहुन्जेल O
, O
बोलबाला O
प्रचण्ड B-PER
कै O
हो O
! O
कंग्रेस B-ORG
+एमाले O
पुच्छर O
हो O
प्रचण्ड B-PER
को O
गोडा O
को O
पानी O
खानु O
अवस्य O
काँग्रेस B-ORG
ले O
5 B-DATE
वर्ष I-DATE
सरकार O
चलाउँछ O
। O
स्थापित O
लाई O
जिताउन O
खोज्ने O
का O
तर्क O
, O
काँग्रेस B-ORG
भित्र O
लडाउँला O
भन्ने O
होला O
त्यै O
भएर O
काँग्रेस B-ORG
ऐ O
ले O
नेकपा B-ORG
काँग्रेस I-ORG
भएको O
होला O
। O
कालान्तर O
मा O
मार्क्स O
र O
कम्युनिस्ट O
विचार O
नै O
काँग्रेस B-ORG
को O
निर्णय O
पनि O
प्रचन्ड B-PER
ले O
गर्न O
थालेछन O
। O
माओवादी B-ORG
सँग O
मिलेर O
काँग्रेस B-ORG
हराउने O
रणनिति O
नै O
गठबन्धन O
थियो O
केरे O
अब O
त O
एमाले B-ORG
पनि O
मिलि O
गो O
, O
धन्यवाद O
राजनेता O
देउवा B-PER
काङ्रेस B-ORG
को O
भोट O
YAML Metadata Warning: empty or missing yaml metadata in repo card (https://huggingface.co/docs/hub/datasets-cards)

DanfeNER - Named Entity Recognition in Nepali Tweets

We have created the largest human annotated Named Entity Recognition (NER) data set for Nepali tweets available to date.

  • DanfeNER covers five named entities - Person, Location, Organization, Event and Date
  • We split the DanfeNER data set into DanfeNER-train and DanfeNER-test. These standard data sets, therefore, become the first benchmark data sets for evaluating tweets Nepali NER systems.
  • We report a comprehensive evaluation of state-of-the-art Transformer models using these data sets.
  • We also discuss the remaining challenges for discovering NEs for Nepali (see our paper below).

Data Set Stats

Data No. Tweets Tokens Avg. Len LOC ORG PER EVT DAT Total Entities
Train 5,366 92,425 17.22 923 782 1,061 34 663 3,463
Test 2,301 39,133 17.00 389 356 444 28 286 1,503
Total 7,667 131,558 17.11 1,312 1,138 1,505 62 949 4,966

Data Format

The DanfeNER data set is divided into train (DanfeNER-train) and test (DanfeNER-test) sets. Each data set has character level as well as token leven annotations. Please read our paper to get more information on this.

Our Results

We used different transformer modles namely: monolingual transformer models as well as multilingual transformer model for our experiment. Monolingual Nepali transformer models are trained from scratch using Nepali text while multilingual models are trained to combine other languages. All of the transformers model used are available on HuggingFace. The different transformer based modles used are as follows:

  • NPVec1-BERT (baseline)
  • NepaliBERT
  • NepBERT
  • DB-BERT
  • BERT-bbmu

Model comparison on DanfeNER-test

Model Pre. Rec. F1
NPVec1-BERT 0.63 0.62 0.63
NepaliBERT 0.72 0.69 0.70
NepBERT 0.71 0.69 0.70
DB-BERT 0.80 0.80 0.80
BERT-bbmu 0.76 0.74 0.75

Performance evaluation of the best performing model (DB-BERT) per named entities:

Model Pre. Rec. F1 Support
PER 0.81 0.77 0.79 444
LOC 0.83 0.86 0.84 389
ORG 0.79 0.79 0.79 356
EVT 0.53 0.29 0.37 28
DAT 0.78 0.84 0.81 286

License

Non-commercial purposes only. For commercial usages, permissions must be taken from the authors and the relevant parties. See the contact address below.

Unless required by applicable law or agreed to in writing, software and data distributed here is on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.

Cite Our Work

If you use the EverestNER data set, please cite our publication:

@inproceedings{niraula2023danfener,
  title={DanfeNER-Named Entity Recognition in Nepali Tweets},
  author={Niraula, Nobal and Chapagain, Jeevan},
  booktitle={The International FLAIRS Conference Proceedings},
  volume={36},
  year={2023}
}

Contact

Feel free to contact nobal @AT nowalab .DOT com and cjeevaniam @AT gmail .DOT com for any inquiries regarding this work.

Acknowledgments

Nepali Shabdakosh - https://nepalishabdakosh.com

Downloads last month
7