progress.md

Progress Report

Pre-3/11 (Zhao):

• Readings: CNN/LSTM, GRU/distance
• Created gru/gru.py
  • Wrote baseline Siamese NN with GRU sentence encoding, Euclidean distance similarity
• Created gru/data_cleaning.py
  • Wrote preprocessing script (from kaggle-quora-dup), saved processed data to data/train_clean.csv
• Downloaded GloVe embeddings to data/, added Data Requirements section to README.md

3/11 (Zhao):

• Tentatively SENT_LEN = 50 while excluding questions less than 4 or greater than 50 in length.
  • Added sentence length visualization and example printing to gru/data_cleaning.py
• Created GRU.md for logging/presentation.
• Began working on Word2Vec model in gru/gru.py

3/12 (Zhao):

• Migrated Word2Vec to word2vecmodel.py
  • Trained 50D, 100D models from training data located in models/

3/12 (Arnav):

• Added Pre-Processed Questions (via textacy)
• Generated Naive Bayes Model
  • Features: Jaccard Similarity, Dissimilar word POS overlap
  • Precision: 0.608
  • Intention: * Non-overlapping content words --> more likely to be non-duplicates
  • Evaluation: * Guided by Jacc. Similarity

3/24 (Arnav):

• Finished SVM Training - updated baseline
• Carried out feature selection by gradually lowering penalty values (NOUN, VERB, ADV most imp.)

3/26 (Zhao):

• Finished nn.py (previous gru.py) - basic framework for training neural nets
• Trained a variety of models, later deleted them due to Keras loading issue with Lambda layers.
  • Collected training/validation accuracy and loss data

4/4 (Zhao):

• Added GloVe word embedding functionality
• Added F1 score computations

4/9 (Zhao):

• Tested Tensorboard, Comet.ml
• Cleaned evaluation