Sentiment Analysis of English Tweets with BERTsent

BERTsent: A finetuned BERT based sentiment classifier for English language tweets.

BERTsent is trained with SemEval 2017 corpus (39k plus tweets) and is based on bertweet-base that was trained on 850M English Tweets (cased) and additional 23M COVID-19 English Tweets (cased). The base model used RoBERTa pre-training procedure.

Output labels:

COVID-19 tweets specific task

Eg., The model distinguishes: "covid" -> neutral sentiment, "I have covid" -> negative sentiment

Cite

If you use BERTsent in your project/research, please cite the following article:
Lamsal, R., Harwood, A., & Read, M. R. (2022). Twitter conversations predict the daily confirmed COVID-19 cases. arXiv preprint arXiv:2206.10471.

@article{lamsal2022twitter,
  title={Twitter conversations predict the daily confirmed COVID-19 cases},
  author={Lamsal, Rabindra and Harwood, Aaron and Read, Maria Rodriguez},
  journal={Applied Soft Computing},
  volume={129},
  pages={109603},
  year={2022},
  publisher={Elsevier}
}

Using the model

Install transformers and emoji, if already not installed:

terminal:
        pip install transformers
        pip install emoji (for converting emoticons or emojis into text)
notebooks (Colab, Kaggle):
        !pip install transformers
        !pip install emoji

Import BERTsent from the transformers library:

from transformers import AutoTokenizer, TFAutoModelForSequenceClassification
  
tokenizer = AutoTokenizer.from_pretrained("rabindralamsal/BERTsent")

model = TFAutoModelForSequenceClassification.from_pretrained("rabindralamsal/BERTsent")

Import TensorFlow and numpy:

import tensorflow as tf
import numpy as np

We have installed and imported everything that's needed for the sentiment analysis. Let's predict sentiment of an example tweet:

example_tweet = "The NEET exams show our Govt in a poor light: unresponsiveness to genuine concerns; admit cards not delivered to aspirants in time; failure to provide centres in towns they reside, thus requiring unnecessary & risky travels. What a disgrace to treat our #Covid warriors like this!"
#this tweet resides on Twitter with an identifier-1435793872588738560
    
input = tokenizer.encode(example_tweet, return_tensors="tf")
output = model.predict(input)[0]
prediction = tf.nn.softmax(output, axis=1).numpy()
sentiment = np.argmax(prediction)
    
print(prediction)
print(sentiment)

Output:

[[0.972672164440155  0.023684727028012276 0.003643065458163619]]
0