nimaafshar's picture
Update README.md
2e2a350

ParsBERT digikala sentiment analysis model fine-tuned on around 600,000 Persian tweets.

How to use

at least you need 650 megabytes of ram and disk in order to load the model. tensorflow, transformers and numpy library

Loading model

import numpy as np
from transformers import AutoTokenizer, TFAutoModelForSequenceClassification
#loading model
tokenizer = AutoTokenizer.from_pretrained("nimaafshar/parsbert-fa-sentiment-twitter")
  
model = TFAutoModelForSequenceClassification.from_pretrained("nimaafshar/parsbert-fa-sentiment-twitter")
classes = ["negative","neutral","positive"]

Using Model


#using model
sequences = [".غذا خیلی افتضاح بود متاسفم برای مدیریت رستورن خیلی بد بود.",
 "خیلی خوشمزده و عالی بود عالی",
"می‌تونم اسمتونو بپرسم؟"
]

for sequence in sequences:
  inputs = tokenizer(sequence, return_tensors="tf")
  classification_logits = model(inputs)[0]
  results = tf.nn.softmax(classification_logits, axis=1).numpy()[0]
  print(classes[np.argmax(results)])
  percentages = np.around(results*100)
  print(percentages)

note that this model is trained on persian corpus and is meant to be used on persian texts too.