Sentiment Roberta finetuned for company reputation analysis
This is a RoBERTa-base model trained on SamEval datasets and fine-tuned with customer tweets. The main task is sentiment analysis with the TweetEval benchmark. The original model can de found here This model is suitable for English. Labels:
- 0 > Negative;
- 1 > Neutral;
- 2 > Positive
Example Pipeline
from transformers import pipeline
sentiment_task = pipeline("sentiment-analysis", model=model_path, tokenizer=model_path)
sentiment_task("Delivery is late!")
Output format: Json
[{'label': 'Negative', 'score': 0.99836}]
Test application here.
Model Details
Model Description
This model is generated to evaluate costumer satisfaction and company reputation.
- Developed by: Anna Chierici
- Language(s) (NLP): English
- Finetuned from model [optional]: cardiffnlp/twitter-roberta-base-sentiment-latest
Training Details
Training Data
SemEval dataset and tweet sent to @AmazonHelp account
Training Procedure
Training Hyperparameters
The following training strategies were implemented:
- validated at the end of each era
- checkpoint saving
- initial learning rate equal to 2e-5
- training/validation batch size of 16
- number of epochs 3
- regularization with weight reduction
- evaluation of the best model at the end of training
- additional metrics accuracy and F1