How Machine Learning Algorithms evolved at Haptik while it's Chatbot catered to 200 million messages
Submitted by krupal Modi (@superkrups) on Friday, 9 June 2017
Full talk for data engineering track
Evolution of automated messaging, which started in 1966 with first Chatbot, ELIZA, has now reached a stage where Chatbots have found there application in several industry domains like personal assistance, customer care, banking, e-commerce, healthcare, etc. With early experiments showing positive results , we have reached a stage where chatbots are no longer merely an application to play around with but have proven their utility in solving real problems. As a result, data scientists need to now figure out how to fuse NLP, conventional machine learning algorithms and deep learning systems into a single dialogue system which can scale easily across datasets from different domains and is capable of digesting training data from real conversations.
During our journey at Haptik, we ended up building and customizing different machine learning modules specifically focused on building Chatbots on narrow domains and targeted at end to end completion of a specific task such as making travel bookings, gift recommendation and ordering, lead generation for different businesses, etc. I would specifically like to share how our machine learning stack grew organically and finally found a stable state containing an ideal mix of simple and complex machine learning algorithms.
In the order of increasing complexity –
Introduction [2-3 mins]
Highlighting different problems that chatbots are solving today with few examples. Introducing why dialogue systems needs to scale and efficiently utilize reseach that happened over last 5 decades.
Keep it simple, start collecting Data [5-7 mins]
How to build a simple system from ground zero which is good enough to go live and helps you collect next million messages.
Analize your conversations, refine the content, make it little more smarter [8 minutes]
Cluster your data and extend your system to use retrieval/classification algorithm to make a bit more intelligent.
When you have enough data[8 mins]
Use complex deep learning models with simpler approaches and utilize every bit of conversational data available with you in most efficient way.
Architecture to stack all the above algorithms[8 mins]
Make sure simpler conversations are catered by simple algorithms and complex ones are in your control while your Chatbot responds fast and accurate.
Challenges [5-7 minutes]
Open challenges existing in the industry and how to foresee/avoid them.
Basic understading of what are Chatbots and what is Machine Learning
I have worked as a Researcher, Engineer and Machine Learning Scientist during different stages of my career. I love to invent, patent, build and architect end to end machine learning solutions to make our life easy. One of my achievements includes creating a chatbot which has seen more than 200 million messages from different domains and is still learning with a long way to go forward. I love to share my learnings with the community by open sourcing and actively participating in Data Science meetups.