Bilingual contrained phrase embedding for MT
Monoligual phrase embedding works with backpropagation
$ time ./phraseEmbedding.py Reading vectors in binary format Read 555587 entries from the binary file Embedding shape = (50,) Cost (0) = 1144.22855732 Iteration : 1 Cost (1) = 571.042721099 Iteration : 2 Cost (2) = 380.347088307 Iteration : 3 Cost (3) = 285.130000339 Iteration : 4 Cost (4) = 228.040957773 Iteration : 5 Cost (5) = 190.001100333 Iteration : 6 Cost (6) = 162.849781705 Iteration : 7 Cost (7) = 142.478591007 Iteration : 8 Cost (8) = 126.64131166 Iteration : 9 Cost (9) = 113.968241981
This experiment had \alpha (Learning rate) = 0.01 \lambda (Regularization parameter) = 0.1 10 iterations of batch backpropagation were performed
TODO: 2. There seem to be some -1s in the input vectors, where are these coming from ? 3. Create a class for phraseEmbedding 4. Do this for two languages 5. Run this for multiple settings of n, \lambda and \alpha 6. Implement fine-tuning (BRAE) mechanism