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.

• Character Level

  • Train: DanfeNER-train-char.txt
  • Test: DanfeNER-test-char.txt

• Token Level

  • Train: DanfeNER-train-bio.txt
  • Test: DanfeNER-test-bio.txt

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