To Example or Not To Example, That is the Natural Language Question - PowerPoint PPT Presentation

1 / 14
About This Presentation
Title:

To Example or Not To Example, That is the Natural Language Question

Description:

... practices around the prompting of natural language ... Offer 2 examples at the initial prompt. Reduce to 1 for re-prompting or looping through NL state ... – PowerPoint PPT presentation

Number of Views:114
Avg rating:3.0/5.0
Slides: 15
Provided by: frederick76
Category:

less

Transcript and Presenter's Notes

Title: To Example or Not To Example, That is the Natural Language Question


1
To Example or Not To Example, That is the
Natural Language Question
  • Frederick Parkinson, PhD, Project Manager, User
    Interface Designer
  • Deborah Rapsinski, Senior User Interface Designer

2
Introduction
  • This paper presents best practices around the
    prompting of natural language recognition states
  • Structure of prompt
  • Use of examples

3
What we will talk about
  • Why we need best practices around Natural
    Language (NL)?
  • What makes designing for NL challenging?
  • What makes for a good NL response?
  • The Nuance best practice for NL
  • Evidence to support this strategy

4
Why we need best practices for NL
  • NL deployments are becoming more common
  • NL deployments incur additional cost, so the
    demand for quick ROI is more acute
  • Best practices allow adopters to achieve ROI
    goals more quickly by allowing designers to
    leverage lessons learned from previous
    deployments

5
Designing for NL is challenging
  • Directed Dialog versus Natural Language
  • Why NL is appealing
  • The Natural Language Paradox
  • Natural Language states are unnatural
  • While callers are encouraged to respond
    naturally, the fact that they are interacting
    with a computer makes the interaction unnatural

6
How Natural Language systems work
  • Two tasks
  • recognition of words in the utterance
  • interpreting the meaning of those words
  • How Natural Language systems are developed
  • trained against corpora of tens of thousands of
    utterances
  • training data is tagged by hand by humans
  • To work well, the recognizer has to pick out a
    few salient words that map to specific caller
    intentions

7
What makes a good Natural Language response
  • A few meaningful words
  • four to ten words long
  • one or two meaningful words
  • Good examples of responses
  • I need to locate a pharmacy, I want to order a
    refill,my power is out, close account
  • Bad examples of responses
  • I need help, inquiry,I need to check about
    amounts of several drugs that I already have
    when it's time to order new ones I don't want to
    order more than I can use

8
The Nuance best practice
  • Provide the caller with examples
  • pro addresses the deer in the headlights some
    callers face
  • con humans dont offer examples
  • Examples provide just enough structure to help
    callers formulate their responses
  • In a few words, please tell me why youre
    calling today. You can say things like I need
    my account balance, or whats the status of my
    order? Okay, go ahead.

9
What makes a good example?
  • Shows callers how to respond 1 statement, 1
    question

Succinct and Unambiguous Fewest number of turns
to destination
Makes use of caller profile If past
due Id like to make payment
arrangements.
10
Best Practice, Continued
  • Rotate examples to teach callers of what can be
    handled
  • Offer 2 examples at the initial prompt
  • Reduce to 1 for re-prompting or looping through
    NL state
  • Balance caller goals with business goals
  • Use examples that will, e.g. encourage callers to
    pay their bill or inform them of new product
    offerings

11
Evidence
  • Usability Testing
  • Inferred success through completed tasks
  • Post-test interviews
  • Experiments
  • No example, 2 statements, 2 questions, 1
    statement and 1 question
  • Deployment Data
  • Fluency, delays
  • Number of words
  • Value of responses

12
Evidence Number of Words Used
No Example
Example Used
13
Evidence Fluency
No Example
Example Used
14
Conclusion
  • Weve presented the Nuance best practice for the
    structure of the Natural Language question
  • Include examples
  • Examples include one statement, one question
  • Examples balance callers goals and business
    goals
  • Weve presented evidence in support of this best
    practice that shows that callers prefer hearing
    examples and that examples improve the quality of
    callers responses to the Natural Language
    question
Write a Comment
User Comments (0)
About PowerShow.com