Abstract
Conversational AI is a system capable of conversing with an interlocutor. Conversational systems, a special variant of question-answering systems, are aware of the context and the conversation history. Translating sequences from one domain to another is an intrinsic characteristic of Neural Machine Translation models (NMT) models and thus, NMT models becomes the natural and the most obvious choice for building high-performing question-answering systems where the task is to convert sequence from question domain to answer domain. GANs are deeply convoluted and highly sophisticated models extremely efficient in generating continuous real-valued data. With some challenges, GANs can be extended to generating textual data. We propose a novel GAN-based architecture that incorporates context-understanding and domain-translation of NMT models with noise understanding and data domain mapping of GANs. The proposed model is a GAN composed of a Sequence-to-Sequence or Seq2Seq network (NMT model) as the generator. We experiment with the idea that a GAN comprised of NMT model will be able to generate meaningful answer sequences and successfully hold a conversation. We have proposed three fundamentally different GAN models that differ from each other based on the loss functions they use, and the architecture of their generator and discriminator. To evaluate the performance of GANs, we have conducted extensive experimentation and tests using key metrics like BLEU scores, F1-scores, and cumulative n-gram scores. One of our best-performing GAN – TextWGAN – using Earth-Mover’s distance as its loss function achieved a BLEU score of 59.7.