metadata
language: fa
license: mit
pipeline_tag: text2text-generation
PersianTextFormalizer
This model is fine-tuned to generate formal text from informal text based on the input provided. It has been fine-tuned on [Mohavere Dataset] (Takalli vahideh, Kalantari, Fateme, Shamsfard, Mehrnoush, Developing an Informal-Formal Persian Corpus, 2022.) using the pretrained model persian-t5-formality-transfer.
Evaluation Metrics
Metric | Basic Model | Base Persian T5 | Our Model |
---|---|---|---|
BLEU-1 | 0.524 | 0.212 | 0.636 |
BLEU-2 | 0.358 | 0.137 | 0.511 |
BLEU-3 | 0.254 | 0.096 | 0.416 |
BLEU-4 | 0.18 | 0.068 | 0.337 |
Bert-Score Precision | 0.671 | 0.537 | 0.797 |
Bert-Score Recall | 0.712 | 0.570 | 0.805 |
Bert-Score F1 Score | 0.690 | 0.549 | 0.800 |
ROUGE-1 F1 Score | 0.553 | - | 0.645 |
ROUGE-2 F1 Score | 0.274 | - | 0.427 |
ROUGE-l F1 Score | 0.522 | - | 0.628 |
Usage
from transformers import (T5ForConditionalGeneration, AutoTokenizer, pipeline)
import torch
model = T5ForConditionalGeneration.from_pretrained('parsi-ai-nlpclass/PersianTextFormalizer')
tokenizer = AutoTokenizer.from_pretrained('parsi-ai-nlpclass/PersianTextFormalizer')
pipe = pipeline(task='text2text-generation', model=model, tokenizer=tokenizer)
def test_model(text):
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model.to(device)
inputs = tokenizer.encode("informal: " + text, return_tensors='pt', max_length=128, truncation=True, padding='max_length')
inputs = inputs.to(device)
outputs = model.generate(inputs, max_length=128, num_beams=4)
print("Output:", tokenizer.decode(outputs[0], skip_special_tokens=True))
text = "به یکی از دوستام میگم که چرا اینکار رو میکنی چرا به فکرت نباید برسه "
print("Original:", text)
test_model(text)
# output: .به یکی از دوستانم می گویم که چرا اینکار را می کنی چرا به فکرت نباید برسد
text = "کجا مخفیش کردی؟"
print("Original:", text)
test_model(text)
# output: کجا او را پنهان کرده ای؟
text = "نمیکشنمون که"
print("Original:", text)
test_model(text)
# output: .ما را که نمیکشند.