The model fine-tunes the bert-base-cased model. It is able to classify text into one of 150 intent classes. The goal is to use this model to classify intents in user input (e.g. emails, customer service, search queries).
The model is trained over 3 epochs on a subset of the training set (2000/7500). This yielded a performance of 77.6% on the test set. By using the full dataset, the accuracy is expected to increase drastically. For reference, the model accuracy improved from 32% to 78% by moving to training on 1000/7500 of the dataset to 2000/7500. Hence, there is significant potential to fine-tune a model for intent classification if the full training set is used. The author did not attempt training on the full dataset due to time constraints.
