BloomBERT is a transformer-based NLP task classifier based on the revised edition of Bloom's Taxonomy.

Bloom's Taxonomy is a set of hierarchical models used in classifying learning outcomes into levels of complexity and specificity. Although mostly employed by educators for curriculum and assessment structuring, BloomBERT takes a novel approach in differentiating the difficulty of a task through classifying productivity related tasks into the cognitive domain of the taxonomy.

BloomBERT can be accessed via an API endpoint or a web application

BloomBERT was built by fine-tuning a DistilBERT model, a lighter version of the original BERT transformer language model developed by Google. It was developed using Tensorflow and the Hugging Face Transformers library, incorporating a sequence classification head (linear layer) on top of the DistilBERT pooled outputs. Utilising the pre-trained DistilBERT model, BloomBERT was trained with a labelled data set curated for the specific task classification on Google Colab.

Training Results:
    EPOCH: 40 
    training accuracy: 0.9820
    validation accuracy: 0.9109

Developed using Streamlit with Python and hosted on Heroku servers through GitHub.
Frontend repository is available here.

Developed using Python to implement FastAPI endpoints. Model was trained using Jupyter Notebook and Tensorflow libraries. Docker used to containerise the application for deployment onto Google Cloud Run.

The API endpoints are currently deployed on Google Cloud.
Note some time may be required for the instance to start up.

GET https://bloom-bert-api-dmkyqqzsta-as.a.run.app

{
  "health_check": "OK", 
  "model_version": "1.0"
}

POST https://bloom-bert-api-dmkyqqzsta-as.a.run.app/predict

{
  "text": "Annotating key points in meeting minutes"
}

{
  "blooms_level": "Understand",
  "probabilities": {
    "Analyse": 0.00078,
    "Apply": 0.00075,
    "Create": 0.00054,
    "Evaluate": 0.00051,
    "Remember": 0.00261,
    "Understand": 0.99481
    }
}

• Naive-Bayes (TF-IDF Vectorizer)
• Naive-Bayes (TF-IDF Vectorizer + SMOTE)
• SVC (TF-IDF Vectorizer)
• SVC (TF-IDF Vectorizer + SMOTE)
• SVC (word2vec, spaCy)
• DistilBERT Transformer model

• Starting with the Naive-Bayes algorithm that is often employed for multiclass classification problems, this model was used as a performance benchmark against other models.

Validation Accuracy: 0.77328

• After observing the presence of data imbalance, Synthetic Minority Oversampling Technique (SMOTE) was implemented in attempts to oversample minority data points to create a balanced dataset and achieve better classification results.
• Using SMOTE successfully improved classification accuracy of the Naive-Bayes Model

Validation Accuracy: 0.81538

• To improve model accuracy, I looked into using a Linear Support Vector Classifier (SVC) for the multi-classification problem.
• SVCs were determined to outperform Naive-Bayes in this specific classification problem with a much higher validation accuracy observed.
• However, the model still fails to generalise well when given inputs of similar semantics.

Validation Accuracy: 0.86721

• Applying SMOTE to this model showed slight improvements in classification accuracy.
• However, it still suffers from the same problems as the above few models.

Validation Accuracy: 0.88421

• To address the problem, I looked into using word2vec (word2vec-google-news-300) in replacement of TF-IDF Vectorizers to extract semantic relations from words within the sentences.
• This model uses spaCy models to tokenise the inputs before feeding the tokens into the word2vec model.
• Each word vector generated from the tokens are then averaged to form a sentence vector input for the SVC model.
• Unexpectedly, there was a significant drop in accuracy compare to the previous model using TF-IDF.

Validation Accuracy: 0.81296

• After doing more research, I approached the problem from another angle using deep learning.
• Transformer models had demonstrated significant improvements over traditional NLP systems, excelling in processing sequential data and understanding language semantics.
• DistilBERT was chosen as the transformer model of choice due to its smaller size and greater speed compared to the original BERT model.
• The pre-trained model was then fine-tuned using the data set that I had developed using the taxonomy.
• It achieved the best accuracy compared to previous models and generalised well to unseen data with similar semantics, providing satisfactory predictions that fit within the taxonomy.
• This was the model chosen for BloomBERT.

Validation Accuracy: 0.91090

Source codes for model development are available under the MIT License. Developed by Ryan Lau Q. F.