Put AI to work
June 26-27, 2017: Training
June 27-29, 2017: Tutorials & Conference
New York, NY

How Amy, an artificial intelligence capable of scheduling meetings, understands human intents

Rakesh Chada (x.ai)
4:00pm4:40pm Thursday, June 29, 2017
Implementing AI
Location: Sutton South/Regent Parlor Level: Beginner

Prerequisite Knowledge

  • A very basic understanding of or interest in natural language processing and generation systems

What you'll learn

  • Gain a high-level overview of x.ai’s production architecture
  • Understand the role and design of an efficient data pipeline for intent classification and the evolution of x.ai’s machine learning techniques
  • Learn a novel and generalizable method for quantifying error tolerance in conversational agents that exploits product tolerance for confusion between similar actions of the agent

Description

Natural language understanding (NLU) is an active area of research within task-oriented conversational agents. The recent emergence of deep learning applications for NLU has dramatically increased the accuracy of virtual assistants and their ability to successfully conduct task-oriented dialogues with people.

Rakesh Chada introduces x.ai’s Amy, an AI assistant that schedules meetings via email by understanding customers preferences and constraints. It is optimized to minimize dialogue ping pong while maximizing success in scheduling a meeting. Amy is not a prototype. It is actively deployed in a production environment and used by paying customers. Since this artificial intelligence interacts with people through emails, NLU performance is central to its design. Incoming emails undergo automatic text information extraction, semantic understanding, and contextual resolution.

Rakesh discusses Amy’s architecture and the various challenges the team faced during its design and shares several machine learning approaches for intent classification. Rakesh concludes by exploring a novel method for error optimization in a conversational agent that exploits customer error tolerance.

Topics include:

  • A high-level overview of x.ai’s production architecture
  • The role and design of an efficient data pipeline for intent classification
  • The evolution of x.ai’s machine learning techniques, from non-deep learning to deep learning
  • Word vectors and the choice between recurrent and convolutional networks for intent classification
  • A novel and generalizable way of quantifying error tolerance in conversational agents that exploits product tolerance for confusion between similar actions of the agent
Photo of Rakesh Chada

Rakesh Chada

x.ai

Rakesh Chada is a data scientist at x.ai building machine learning systems to understand human intents from emails. He holds a master’s degree in computer science from Stony Brook University with focus on machine learning and natural language processing, where he worked on question-answering systems, Wikipedia graph mining, topic modeling, and the like under Steven Skiena.