Session 3 - Preprocessing, Transformation, and Classification Models

Solution 09 - Sentiment Predictor (Bonus)

Learning objective: In this exercise, you will further practice performing classification on an IMDB data set.

Workflow description: This workflow predicts the sentiment of movie reviews using two different supervised approaches: training a ML model from scratch, and using transfer learning.

You’ll find the instructions for the exercises in the yellow annotations.

Activity 1. Supervised learning: Build a sentiment predictor from scratch

1. Create document objects out of movie reviews.

2. Preprocess documents.

   • Remove numbers, characters, punctuation, stop words, etc.

   • Convert text to lower case.

   • Perform stemming.

3. Shuffle the dataset with the Shuffle node and split it into training (70%) and test set (30%) with stratified sampling on the target column.

4. Create a bag of words and compute relative term frequency for the training and test set separately.

5. Create document vectors for the training and test set.

6. Extract sentiment class label using the Category To Class node.

7. Train a Decision Tree model and score it using the Scorer and the ROC Curve nodes.

   ---

Your Solution

Activity 2. Supervised learning: Build a sentiment predictor with transfer learning

1. Retain only movie reviews and target column.

2. Shuffle the dataset and split it into training (70%) and test set (30%) with stratified sampling on the target column.

3. Lowercase text reviews.

4. Use the BERT Model Selector node to download the pretrained bert_en_uncased_L-12_H-768_A-12 from TensorFlow Hub.

5. Fine-tune the pretrained model with the BERT Classification Learner node. Set max sequence length to 128, no. of epochs to 2, batch size to 128, validation to 20, optimizer to Adam and learning rate to 0,00001.

6. Apply the fine-tuned model with the BERT Predictor node and score it.

   ---

Your Solution