BaurTokenov/NullitOut

NullitOut

Plan:

• Look at the implementation code (Tue)
  • https://github.com/shauli-ravfogel/nullspace_projection
• Find dataset format (Tue)
• Preprocess our data (format it) (Fri)
  • https://www.kaggle.com/assemzhunis/nullitout/
• Run the null it out algo (Sun)
  • https://colab.research.google.com/drive/1zB8SV-c6t-PGindgnZR0P8KO_wpFOVg_?usp=sharing

1. Convert our datase to embeddings: (get_embeddings_based_dataset(comment_text))

• Fasttext
• Bert

2. Train toxicity classifier

• Profession classifier (start with logistic regression) Y_dev is 'toxicity' (0,1)
• Savethe performance (Acc)

3. Build ethnicity classifier (Gender classifier) and run NIO

• get_projection_matrix()
• get accurcy (we need to reduce this)

4. Run toxicity classifier on our new embeddings

• get performance (we want this to be somewhat the same as before)

• Plot the results (Sun)

• Analysis report (Sun)

• Presentation (Mon)