SpeechCycle Confidential Confidential 1 Optimizing Natural Language Interfaces: No Data Like More Data SpeechTEK New York, 2007 Jonathan Bloom & Roberto.

Slides:



Advertisements
Similar presentations
What is Science? Science refers to a method of learning about the natural world, as well as to the knowledge gained through that process. Scientific Inquiry.
Advertisements

Chapter 14: Usability testing and field studies
The Social Scientific Method An Introduction to Social Science Research Methodology.
Tuning Jenny Burr August Discussion Topics What is tuning? What is the process of tuning?
Dialogue Policy Optimisation
1 Profit from usage data analytics: Recent trends in gathering and analyzing IVR usage data Vasudeva Akula, Convergys Corporation 08/08/2006.
Scientific Method Chapter 1.
Managing Complexity: 3rd Generation Speech Applications Roberto Pieraccini August 7, 2006.
Chapter 14: Usability testing and field studies. 2 FJK User-Centered Design and Development Instructor: Franz J. Kurfess Computer Science Dept.
Implementation of Computer Simulation Software in Learning Low-Level Computer Language: A Case Study Johnny Chan Chi Hown.
PPA Advisory Board Meeting, May 12, 2006 Assessment Summary.
VoiceXML Application Systems Team By: Shawn Ramdass, Saji Abraham & Billy Santamorena.
Introduction to Educational Research
Chocolate Bar! luqili. Milestone 3 Speed 11% of final mark 7%: path quality and speed –Some cleverness required for full marks –Implement some A* techniques.
Beyond Usability: Measuring Speech Application Success Silke Witt-Ehsani, PhD VP, VUI Design Center TuVox.
How do Scientists Think?
Science, Society, and You
1 Validation & Verification Chapter VALIDATION & VERIFICATION Very Difficult Very Important Conceptually distinct, but performed simultaneously.
1 High Resolution Statistical Natural Language Understanding: Tools, Processes, and Issues. Roberto Pieraccini SpeechCycle
2008 Science Summer School A quick report – Choosing Science Survey.
Convergys Confidential and Proprietary Automating Redelivery for the USPS Kristie Goss Convergys Corporation
Holistic Usability Measure VUI Design: Philosophy and Practice.
Speech and Language Processing Chapter 24 of SLP (part 3) Dialogue and Conversational Agents.
CHAPTER 1 Scientific Method. Scientific Method (yes, copy these steps!) The scientific method is a series of steps used to solve problems. Steps: 1. State.
What is Science? Science is a system of knowledge based on facts and principles.
© 2008 SpeechCycle, Inc. More Than Call Steering Managing Dynamic Contextual Complexity Phillip Hunter VP, Application Design/Development SpeechTek NYC.
Lesson Designing Samples. Knowledge Objectives Define population and sample. Explain how sampling differs from a census. Explain what is meant by.
The Science of Biology Chapter 1 Biology Ms. Haut.
LEVEL 3 I can identify differences and similarities or changes in different scientific ideas. I can suggest solutions to problems and build models to.
Learning Automata based Approach to Model Dialogue Strategy in Spoken Dialogue System: A Performance Evaluation G.Kumaravelan Pondicherry University, Karaikal.
Environmental Science Chapter 2 – Scientific Tools Test Review
What do physical scientists study to learn about the world?
Unit: Science & Technology Lesson #3 Scientific Inquiry Essential Question: What is scientific inquiry? How do you design and conduct an experiment? What.
1 CS 224S W2006 CS 224S LING 281 Speech Recognition and Synthesis Lecture 15: Dialogue and Conversational Agents (III) Dan Jurafsky.
Process & Inquiry Skills: Scientists in EC Classrooms CDAEYC Annual Meeting May 19, 2015 Albany, New York.
ICCS WSES BOF Discussion. Possible Topics Scientific workflows and Grid infrastructure Utilization of computing resources in scientific workflows; Virtual.
Nursing Research as the Basis of Nursing. Importance of Nursing Research Nurses ask questions aimed at gaining new knowledge to improve pt. care Nurses.
 There isn’t a single scientific method, but there is a style of investigation that can be called scientific methodology.  There are 5 main parts that.
Chapter 1 Scientific Method. Observing is when you notice and describe events or processes in a careful, orderly way. (My cell phone won't work...what's.
Review of the Scientific Method Chapter 1. Scientific Method – –Organized, logical approach to scientific research. Not a list of rules, but a general.
1.2 Using a Scientific Approach. 1. Make an Observation information obtained by using your five senses. Hearing, seeing, smelling, touching, or tasting.
AutoTutor Benjamin Kempe Tutoring Research Group, University of Memphis
The Language of Science.  Hypothesis: a prediction that can be tested; an educated guess base on observations and prior knowledge  Theory: a well tested.
Chapter 1.1 – What is Science?. State and explain the goals of science. Describe the steps used in the scientific method. Daily Objectives.
1.3 Scientific Thinking and Processes KEY CONCEPT Scientific Method Science is a way of thinking, questioning, and gathering evidence.
Teaching & Inquiry I: Fundamentals of Teaching Through Inquiry Todd Twyman 9/18(20)
Scientific Method Chapter 1-1. What is Science?  Science – organized way of gathering and analyzing evidence about the natural world  Described as a.
What is Science? SECTION 1.1. What Is Science and Is Not  Scientific ideas are open to testing, discussion, and revision  Science is an organize way.
Qualities of an Aligned Lesson Aligning Content and Process.
Scientific Method 1.Observe 2.Ask a question 3.Form a hypothesis 4.Test hypothesis (experiment) 5.Record and analyze data 6.Form a conclusion 7.Repeat.
Conducting Research Psychology, like chemistry and biology, is an experimental science, assumptions must be supported by scientific evidence. It is not.
Evidence Based Practice & Research in Nursing Level 8, Academic Year (AY) 1434—1435 H Vanessa B. Varona, RN, MAN.
Introduction to Consumer Behavior and Marketing Strategy
Intro to AP Statistics and Exam
Scientific Method and Experiment Additional Terms
Research Methods for Computer Science
Scientific Inquiry with Gummy Bears
Scientific Method notes and flow chart
Title of your science project
What is Physical Science?
SCSH3. Students will identify and investigate problems scientifically
Answer the following questions
Reinforcement Learning with Partially Known World Dynamics
Chapter 1.1 – What is Science?
Unit: Science and Technology
Scientific Inquiry Ch1: Lesson 3
What processes do scientists use when they perform scientific investigations? Chapter Introduction.
Scientific Inquiry.
Presentation transcript:

SpeechCycle Confidential Confidential 1 Optimizing Natural Language Interfaces: No Data Like More Data SpeechTEK New York, 2007 Jonathan Bloom & Roberto Pieraccini

SpeechCycle Confidential Confidential 2 SpeechCycle Confidential Just as speech scientists crunch data to optimize a speech recognizer, speech companies need to crunch data to optimize a call at the dialog level as well. Voice user interface (VUI) design needs to be based on quantitative data as much as possible. We will provide an example of quantitative research in the area of VUI design. We will ask, “What is the future of this research-based VUI approach?” Executive Summary

SpeechCycle Confidential Confidential 3 SpeechCycle Confidential VUI = RELIGION + SCIENCE

SpeechCycle Confidential Confidential 4 SpeechCycle Confidential Where there is data, we use it. At those times, it is a science. All else is faith-based. Religion is a fine thing. But our customers do not pay us for theology lessons. They pay us to save them $$$ and keep their customers happy. VUI = RELIGION + SCIENCE

SpeechCycle Confidential Confidential 5 As much as possible, speech companies need to collect data in order to optimize their voice user interfaces at every level – and in every facet - of the interaction, from recognition accuracy to prompt wording to dialog structure. In doing so, VUI will become less anecdotal and more scientific. The Point

SpeechCycle Confidential Confidential 6 One needs the ability to compare two or more versions of an application… …running at the same time. …taking calls from the same population of callers. …taking calls from that population at random. …gathering data at a fast enough pace to meet customer deadlines. Requirements for Exploration

SpeechCycle Confidential Confidential 7 Designing Exploration Alternatives

SpeechCycle Confidential Confidential 8 Example “All of our agents are currently helping other customers. Let’s get started with our automated internet troubleshooter, so you don’t have to wait. [chime] To begin, briefly describe the problem, saying something like “I can’t connect to the internet”, or you can say “What are my choices”. “All of our agents are currently helping other customers. Let’s get started with our automated internet troubleshooter, so you don’t have to wait. [chime] Are you calling because you’ve lost your internet connection? [pause] Please say yes or no.”

SpeechCycle Confidential Confidential 9 Example RESULTS INTRO followed by SLM = 20.4% automation rate INTRO followed by YES/NO question = 22.6% automation rate (CHI square) Statistically significant at.05 level

SpeechCycle Confidential Confidential 10 Summary (Up To Now) Strong opinions and qualitative usability tests should not be the only sources of VUI knowledge. With the right tools, continual access to data, and with enough data, speech companies can make dialog experimentation a regular part of the product lifecycle. At this point, other than the randomization script in the call flow, a lot of this process is manual, so now we need to ask, “How can we make this process more automatic?” What’s next?

SpeechCycle Confidential Confidential 11 Reinforcement Learning theory A finite choice of actions What the system does A finite or infinite set of interaction states Identify the factors that can influence the choice of the best action Policy Choose the action based on current state of the interaction Maximization of a return function Reinforce the choice of the actions that provide a positive return at any particular state The process converges to a locally optimal policy

SpeechCycle Confidential Confidential 12 Reinforcement Learning in dialog research Research has proven that reinforcement learning can be used for automating the design of dialog systems (Markov Decision Processes) E. Levin and R. Pieraccini (AT&T), 1997 (Partially Observable Markov Decision Processes) S. Young (Univ. Cambridge, 2004) Unfortunately, full dialog learning and optimization requires a lot of interactions Academic research uses simulated users Restricting the optimization to a small number of reasonable “competing designs” Alternative competing designs at different points in the application Exploration and Exploitation principle

SpeechCycle Confidential Confidential 13 Conclusions VUI design—from religion to science Use data to validate and chose optimal design among competing alternatives Reinforcement Learning—from theory to practice Use data to optimize applications while they interact