File size: 5,820 Bytes
4e4157d dd598b6 4e4157d dd598b6 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 |
---
language:
- bn
metrics:
- f1
pipeline_tag: token-classification
---
# Bangla-Person-Name-Extractor
This repository contains the implementation of a Bangla Person Name Extractor model which is able to extract Person name entities from a given sentence. We approached it as a token classification task i.e. tagging each token with either a Person's name or not. We leveraged the [BanglaBERT](http://https://github.com/csebuetnlp/banglabert) model for our task, finetuning it for a binary classification task using a custom-prepare dataset. We have deployed the model into huggingface for easier access and use case.
# How to use it?
[This Notebook](https://github.com/MBMMurad/Bangla-Person-Name-Extractor/blob/main/Inference_template.ipynb) contains the required Inference Template on a sentence.
<br></br>
You can also directly infer using the following code snippet. Just change the sentence.
```
from transformers import AutoModelForPreTraining, AutoTokenizer,AutoModelForTokenClassification #!pip install transformers==4.30.2
from normalizer import normalize #pip install git+https://github.com/csebuetnlp/normalizer
import torch #pip install torch
import numpy as np #!pip install numpy==1.23.5
model = AutoModelForTokenClassification.from_pretrained("MBMMurad/BanglaBERT_Person_Name_Extractor")
tokenizer = AutoTokenizer.from_pretrained("MBMMurad/BanglaBERT_Person_Name_Extractor")
def inference_fn(sentence):
sentence = normalize(sentence)
tokens = tokenizer.tokenize(sentence)
inputs = tokenizer.encode(sentence,return_tensors="pt")
outputs = model(inputs).logits
predictions = torch.argmax(outputs[0],axis=1)[1:-1].numpy()
idxs = np.where(predictions==1)
return np.array(tokens)[idxs]
sentence = "আব্দুর রহিম নামের কাস্টমারকে একশ টাকা বাকি দিলাম।"
pred = inference_fn(sentence)
print(f"Input Sentence : {sentence}")
print(f"Person Name Entities : {pred}")
sentence = "ইঞ্জিনিয়ার্স ইনস্টিটিউশন চট্টগ্রামের সাবেক সভাপতি প্রকৌশলী দেলোয়ার হোসেন মজুমদার প্রথম আলোকে বলেন, 'সংকট নিরসনে বর্তমান খালগুলোকে পূর্ণ প্রবাহে ফিরিয়ে আনার পাশাপাশি নতুন তিনটি খাল খনন জরুরি।'"
pred = inference_fn(sentence)
print(f"Input Sentence : {sentence}")
print(f"Person Name Entities : {pred}")
sentence = "দলীয় নেতারা তাঁর বাসভবনে যেতে চাইলে আটক হন।"
pred = inference_fn(sentence)
print(f"Input Sentence : {sentence}")
print(f"Person Name Entities : {pred}")
```
**Output:**
```
Input Sentence : আব্দুর রহিম নামের কাস্টমারকে একশ টাকা বাকি দিলাম।
Person Name Entities : ['আব্দুর' 'রহিম']
Input Sentence : ইঞ্জিনিয়ার্স ইনস্টিটিউশন চট্টগ্রামের সাবেক সভাপতি প্রকৌশলী দেলোয়ার হোসেন মজুমদার প্রথম আলোকে বলেন, 'সংকট নিরসনে বর্তমান খালগুলোকে পূর্ণ প্রবাহে ফিরিয়ে আনার পাশাপাশি নতুন তিনটি খাল খনন জরুরি।'
Person Name Entities : ['দেলোয়ার' 'হোসেন' 'মজুমদার']
Input Sentence : দলীয় নেতারা তাঁর বাসভবনে যেতে চাইলে আটক হন।
Person Name Entities : []
```
# Datasets
We used two datasets to train and evaluate our pipeline.
1. [Bengali-NER/annotated data at master · Rifat1493/Bengali-NER](http://https://github.com/Rifat1493/Bengali-NER/tree/master/annotated%20data)
2. [banglakit/bengali-ner-data](http://https://raw.githubusercontent.com/banglakit/bengali-ner-data/master/main.jsonl)
The annotation formats for both datasets were quite different, so we had to preprocess both of them before merging them. Please refer to [this notebook](https://github.com/MBMMurad/Bangla-Person-Name-Extractor/blob/main/prepare-dataset.ipynb) for preparing the dataset as required.
# Training and Evaluation
We treated this problem as a token classification task.So it seemed perfect to finetune the BanglaBERT model for our purpose. [BanglaBERT ](https://huggingface.co/csebuetnlp/banglabert)is an [ELECTRA](https://openreview.net/pdf?id=r1xMH1BtvB) discriminator model pretrained with the Replaced Token Detection (RTD) objective. Finetuned models using this checkpoint achieve state-of-the-art results on many of the NLP tasks in bengali.
We mainly finetuned two checkpoints of BanglaBERT.
1. [BanglaBERT](https://huggingface.co/csebuetnlp/banglabert)
2. [BanglaEERT small](https://huggingface.co/csebuetnlp/banglabert_small)
BanglaBERT performed better than BanglaBERT small ( 83% F1 score vs 79% F1 score on the test set) .
Please refer to [this notebook](https://github.com/MBMMurad/Bangla-Person-Name-Extractor/blob/main/Training%20Notebook%20%3A%20Person%20Name%20Extractor%20using%20BanglaBERT.ipynb) to see the training process.
**Quantitative results**
Please refer to [this notebook](https://github.com/MBMMurad/Bangla-Person-Name-Extractor/blob/main/Inference%20and%20Evaluation%20Notebook.ipynb) to see the evaluation process.
<br></br>
|