We look in to the virtual assistant robotic system. But most of the chatbot uses Dialogflow or Rasa uses just simple strategy to response. We need an assistant that constantly fine-tune based on certain set of policies. It can be sentiment or fake news or any function. An actively learning strategy takes time. To make things faster we added PyTorch based QnA model that is trained on Stanford Question Answering Dataset (SQuAD) and various COVID-19 articles.
What it does
We upload PDF files of various articles, news related to the pandemic in the backend, the PyTorch model will do a semantic search and find the best answer. The question and answers are saved as source and target file to train our Seq2Seq model. The model gets fine tuned with pretrained sentimental anlalysis model using RL which will be deployed in the edge. The model get's trained in backend in shadow mode, untill best result achieved. Mean time our PyTorch QnA model will helps the user. When the model achived desired accuracy, it will deployed in the edge.
How we built it
We built it using Python, PyTorch, TensorFlow
Challenges we ran into
Optimizing the code for Tensorflow 1.x is complicated, since most of the modules are depreciated and required a lot of debugging
Accomplishments that we're proud of
The agent learned in an active way
What we learned
Built a succesful model
What's next for seq2seq RL chatbot with PyTorch QnA
Deploying as a web application that uses powerful ML instance severs