Continuous Improvement of Chatbots


Human in the Loop for Improving Modeling

In order to bring more trust while deploying deep learning based models in conversational systems, we are exploring ways in which systems built using these technology could be deployed in a controlled environment and could be improved as they are used in that environment. One natural place for achieving this in the case of customer care is to use human agents in contact centers. We are exploring deep learning based methods that could be used to generate responses given customer queries as well recommend documents or small textual chunks that could be used to solve the given customer problems. We are also exploring incremental learning methods that could use the interactions done by human agents on these recommendations to further improve the systems.

A key challenge in building a bot is to handle unrecognized requests -requests that are not currently being captured with high confidence by an intent classifier. In order to address this problem, we have been working on the problem of intent recommendations. We use a fast clustering algorithm developed in-house which leverages sentence embeddings from large transformer based models to do unsupervised clustering. The clusters are then displayed to Tanya (human in he loop) who uses the recommendations to add new intents or add missing examples to existing intents. This helps the bot designer to improve the recognition rate of the bot.

Bootstrapping and continuous improvement of chatbots

For building chat-bots using frameworks such as Watson Assistant and DialogFlow, a dialog designer spends enormous time in figuring out what problems are going to be asked by the bot and how should they be responded and then modeling them in the form of intents, entities and dialog flow. We have been exploring how we could use past human to human conversations and learn these model artifacts automatically. We have also been exploring on explainable deep learning techniques for conversation modeling. 

Speech Interface over chatbots

With the advances in conversational AI systems, the natural extension of using such systems is via a speech interface which adds to the accessibility of the chatbots. While traditional Speech-to-Text (STT) and Text-to-Speech (TTS) systems have improved tremendously, they do have recognition errors in the constrained domains modelled by the chatbots. In this project, we leverage the chatbot artifacts such as intents, entities and dialog states to provide effective customization to the STT models so that their in-domain performance improves.

Our primary focus is to come up with new techniques that would enable effective adaptation of existing STT and TTS models with small amount of domain specific data without impacting the inference time. We explore new model architectures for the STT that can adapt quickly and more effectively to dynamic chatbot information