TODO (ALEX):
-
Fix embedding freeze ability https://discuss.pytorch.org/t/freeze-the-learnable-parameters-of-resnet-and-attach-it-to-a-new-network/949/11
-
Experiment with Vocab Size
-
Experiment with word adjustments (numbers -> N, removing stop_words, UNK heuristics, word roots)
-
Remove all words in common
-
Add Batchnorm
-
LSTM multilayer MLP
-
LSTMs with: cosine similarity, "angle", "distance"
-
LSTM with hand crafted features concatenated on top
-
Recurrent CNN Model
TODO(Abhishek):
- Review code.
- Check the hyperparameters for the experiments.
TODO(Cipta):
- Finish autoencoder experiments
- Integrate with Alex's models
TODO (ALL):
- Ensemble Best Models
- Linear Interpolation
- Hand Crafted Features concatenated with model predictions and run through model
Hand Crafted Features:
- n words q1, n words q2
- difference in word count
- pct word similarity
- number of common words
- cosine similarity of avrage of word2vex embeddings
- pos tags
- Start with same quesiton word
Evaluation:
- Accuracy
- Precision
- Recall
- F1