The implementation differs from the original paper in these ways:
- no lexicons
- Nadam optimizer used instead of SGD
- Parameters: LSTM cell size of 200 (vs 275), dropout of 0.5 (vs 0.68)
The implementation achieves a test F1 score of ~86 with 30 epochs. Increase the number of epochs to 80 reach an F1 over 90.
Indian Culinary Dataset GloVe vector representation from Jeffrey Pennington, Richard Socher, and Christopher D. Manning. 2014. See https://nlp.stanford.edu/projects/glove/
1) numpy
2) Keras
3) Tensorflow
4) Stanford GloVE embeddings