Create README.md
Browse files
README.md
ADDED
@@ -0,0 +1,54 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
## ParsTwiNER: Transformer-based Model for Named Entity Recognition at Informal Persian
|
2 |
+
An open, broad-coverage corpus and model for informal Persian named entity recognition collected from Twitter.
|
3 |
+
Paper presenting ParsTwiNER: [2021.wnut-1.16](https://aclanthology.org/2021.wnut-1.16/)
|
4 |
+
---
|
5 |
+
## Results
|
6 |
+
The following table summarizes the F1 score on our corpus obtained by ParsTwiNER as compared to ParsBERT as a SoTA for Persian NER.
|
7 |
+
### Named Entity Recognition on Our Corpus
|
8 |
+
| Entity Type | ParsTwiNER F1 | ParsBert F1 |
|
9 |
+
|:-----------:|:-------------:|:--------------:|
|
10 |
+
| PER | 91 | 80 |
|
11 |
+
| LOC | 82 | 68 |
|
12 |
+
| ORG | 69 | 55 |
|
13 |
+
| EVE | 41 | 12 |
|
14 |
+
| POG | 85 | - |
|
15 |
+
| NAT | 82.3 | - |
|
16 |
+
| Total | 81.5 | 69.5 |
|
17 |
+
|
18 |
+
## How to use
|
19 |
+
### TensorFlow 2.0
|
20 |
+
```python
|
21 |
+
from transformers import TFAutoModelForTokenClassification
|
22 |
+
tokenizer = AutoTokenizer.from_pretrained("overfit/twiner-bert-base-mtl")
|
23 |
+
model = TFAutoModelForTokenClassification.from_pretrained("overfit/twiner-bert-base-mtl")
|
24 |
+
twiner_mtl = pipeline('ner', model=model, tokenizer=tokenizer, ignore_labels=[])
|
25 |
+
```
|
26 |
+
## Cite
|
27 |
+
Please cite the following paper in your publication if you are using [ParsTwiNER](https://aclanthology.org/2021.wnut-1.16/) in your research:
|
28 |
+
```markdown
|
29 |
+
@inproceedings{aghajani-etal-2021-parstwiner,
|
30 |
+
title = "{P}ars{T}wi{NER}: A Corpus for Named Entity Recognition at Informal {P}ersian",
|
31 |
+
author = "Aghajani, MohammadMahdi and
|
32 |
+
Badri, AliAkbar and
|
33 |
+
Beigy, Hamid",
|
34 |
+
booktitle = "Proceedings of the Seventh Workshop on Noisy User-generated Text (W-NUT 2021)",
|
35 |
+
month = nov,
|
36 |
+
year = "2021",
|
37 |
+
address = "Online",
|
38 |
+
publisher = "Association for Computational Linguistics",
|
39 |
+
url = "https://aclanthology.org/2021.wnut-1.16",
|
40 |
+
pages = "131--136",
|
41 |
+
abstract = "As a result of unstructured sentences and some misspellings and errors, finding named entities in a noisy environment such as social media takes much more effort. ParsTwiNER contains about 250k tokens, based on standard instructions like MUC-6 or CoNLL 2003, gathered from Persian Twitter. Using Cohen{'}s Kappa coefficient, the consistency of annotators is 0.95, a high score. In this study, we demonstrate that some state-of-the-art models degrade on these corpora, and trained a new model using parallel transfer learning based on the BERT architecture. Experimental results show that the model works well in informal Persian as well as in formal Persian.",
|
42 |
+
}
|
43 |
+
```
|
44 |
+
## Acknowledgments
|
45 |
+
The authors would like to thank Dr. Momtazi for her support. Furthermore, we would like to acknowledge the accompaniment provided by Mohammad Mahdi Samiei and Abbas Maazallahi.
|
46 |
+
## Contributors
|
47 |
+
- Mohammad Mahdi Aghajani: [Linkedin](https://www.linkedin.com/in/mohammadmahdi-aghajani-821843147/), [Github](https://github.com/mmaghajani)
|
48 |
+
- Ali Akbar Badri: [Linkedin](https://www.linkedin.com/in/aliakbarbadri/), [Github](https://github.com/AliAkbarBadri)
|
49 |
+
- Dr. Hamid Beigy: [Linkedin](https://www.linkedin.com/in/hamid-beigy-8982604b/)
|
50 |
+
- Overfit Team: [Github](https://github.com/overfit-ir), [Telegram](https://t.me/nlp_stuff)
|
51 |
+
|
52 |
+
## Releases
|
53 |
+
### Release v1.0.0 (Aug 01, 2021)
|
54 |
+
This is the first version of our ParsTwiNER.
|