Authors: Ananya (12040180), Vaibhav (12041650)
This Repo contains solution for DS605 (Deep Learning for low-resource NLP) assignment. This assignment basically revolves around building unsupervised machine translation system using discrete and continuous variance autoencoder models given certain subtasks to be taken into account.
Assignment (50 Marks) What steps would you follow to build an unsupervised machine translation system using discrete and continuous variance autoencoder models? Additionally, deal with the subsequent sub-issues:
(i) What kind of encoding will this machine translation system utilize, and why?
(ii) What kind of encoder-decoder model would be employed for this machine translation system, and why?
(iii) What is the latent variable's expected loss value, and how will you define it? Hint: Although JS-divergence and Energy distance are both smooth functions and differentiable, KL-divergence is not.
(iv) Most text-VAE-based models do not support standard training methods such as backpropagation. Suggest some suitable approaches for training the model.
Bonus Point (50% of the assignment marks (25 Marks)): Design an unsupervised MT system that addresses the above-mentioned problems (and sub-problems) for the English-Hindi dataset. The submitted code will be evaluated based on metrics of performance (BLEU and TER).
- Solution is in this pdf of the repo. Please download the file, if unable to view on github
- Code is in this file of the repo
-
He, J. X. W., & Berg-Kirkpatrick, T. (2020). A Probabilistic Formulation of Unsupervised Text Style Transfer. ICLR.
-
Higgins, I., Matthey, L., Pal, A., Burgess, C., Glorot, X., Botvinick, M., ... & Lerchner, A. (2017). Beta-VAE: Learning basic visual concepts with a constrained variational framework. ICLR.
-
Bowman, S. R., Vilnis, L., Vinyals, O., Dai, A. M., Jozefowicz, R., & Bengio, S. (2016). Generating sentences from a continuous space. arXiv preprint arXiv:1511.06349.
-
Kingma, D. P., Salimans, T., & Welling, M. (2016). Improving variational inference with inverse autoregressive flow. NIPS.