Sentiment analysis
Use transformers to predict customer sentiment
CamemBERTA history of CamemBERT !
As a continuation of my work-study program at My-Serious-Game, I carried out a Deep Learning project using the CamemBERT library (Transformers) on satisfaction survey responses collected at the end of modules on their LMS. Learners have the opportunity to leave comments reflecting their experience with the module.
First, with the help of a subject-matter expert, I labeled a dataset (as positive, negative, or neutral) based on customer satisfaction feedback from EdMill. A total of 1,400 entries were labeled.
Next, I prepared the data by cleaning and tokenizing the text.
I then adapted the pre-trained model for sentiment analysis by modifying the layers (Dense, etc.) for classification.
After that, I divided the data into batches and configured training parameters such as the learning rate, number of epochs, and optimizer. For sentiment analysis, a moderate number of epochs (e.g., 2 to 5) is often sufficient.
The final step was training the model. I monitored the loss and accuracy during training to avoid overfitting and to ensure that the model was learning effectively.
Screenshots :
