Adversarial Training of Text for robust Domain Adaptation using Deep Learning
- 0 Collaborators
In this project we adopt Adversarial Training for textual data in different domains, which is otherwise used extensively for images. A semi-supervised approach will be used to handle the dearth of labelled data in any domain. We explore SOTA, BERT features and Deep Learning architectures. ...learn more
Project status: Concept
Groups
DeepLearning
Intel Technologies
Other
Overview / Usage
Domain Adaptation aims to learn the underlying data distribution of a data-set from a certain domain so that the common features are learnt that can be used while testing on data-set from some other domain. This helps in tackling issues of limited data in various domains which is very much evident in health care, legal documents, among many others. Also, supervised algorithms require labelled data whose curation is intensive, time-consuming and costly process that involves humans annotating them.
Very recently Adversarial Training to tackle the above issues that was used extensively in the image domain. This technique can be used for semi-supervised and unsupervised implementation with the help of a Discriminator network and Gradient Reversal Training. In this project we explore this method in addition to using state-of the art Natural Language Processing models like BERT and architectures like LSTMs and CNNs, to enhance the performance of the model.
Technologies Used
PyTorch
Python
BERT
LSTMs
Convolutional Neural Network