Transcript No Slide Title

CAREER: Intelligent Generation of
Text and Information Graphics*
• Motivation:
• vital technical information involving scientific or
medical arguments may be difficult for lay person to
grasp
• Proposal:
• use AI to help technical experts produce “user-friendly”
arguments in text and/or graphics
• use HCI methods to ensure effectiveness
• build demonstration system (GenIE) for genetic
counselors
*This material is based upon work supported by the National Science Foundation under Grant No. 0132821. Any opinions, findings, and conclusions or
recommendations expressed in this material are those of the author and do not necessarily reflect the views of the National Science Foundation.
The Genetic Counselor
• Meets with clients
– Informational/educational role
• Explanation of diagnosis of genetic condition
• Explanation of inheritance risks
• General explanation of background on genetics
– Counseling role
• Writes summary letter (1-2 pages) for client
Client Issues
• Complex Subject
–
–
–
–
–
probability and statistics
hypothetical outcomes
causality
scientific and medical terminology
diagrams may help
• Emotional Distress
• Reader’s ability to comprehend (innumeracy)
• Rapidly changing information
GenIE: Genetics Intelligent Editor
Goals:
• NOT to replace human counselor
• Reduce counselor’s effort
– artificial intelligence creates 1st draft of letter
– human counselor may revise or reject GenIE’s draft
• Design online presentation; client benefits:
– supplementary graphics and animation
– links to other resources
– automatic updates
Multiple Research Methods
Goal: Evaluate presentation techniques
under controlled conditions in lab
HCI
Experiments
Computational
Model Building
Corpus
Analysis
Goal: develop widely
applicable computational
(AI) techniques for
generating arguments
Goal: study corpus to understand how
human authors communicate technical
arguments
GenIE
Goal: concrete
implementation of ideas
for demonstration and
evaluation
Research Methods: HCI Experiments
• Before: Formally evaluate effectiveness of
communication techniques before computational
models created, e.g.:
– How does layout of document affect comprehension of arguments?
– What types of information to present in text, in graphics, or both?
– Graphical depiction of argument structure
• After: Evaluate communicative effectiveness of
presentations created by GenIE
– ablation experiments to identify which factors contribute or detract
from communicative effectiveness
Research Methods: Corpus Analysis
• Corpus Acquisition (text and graphics)
– genetic counseling summary letters, client education
documents (print and web)
• Qualitative Analysis
– types of information & graphic techniques
– analysis of argumentation (ex. predictive, diagnostic,
value-based, Toulmin-style, dialectical)
• Computational Linguistics Analysis
– develop coding scheme with intercoder reliability
– manually encode corpus
– manual and automated discovery of communication
techniques “evolved” by human authors
Research Methods: Computational
Models
Develop AI methods to
– represent the underlying scientific arguments and
reasoning of the experts
– predict the audience’s potential problems in
understanding, e.g.,
• complexity of causal explanation
• emotionally disturbing information
– reason about content (both text & graphics), organization,
and layout to avoid predicted problems
– generate text and graphics based on above
Analysis of Argumentation in Corpus
• Argumentation: discourse that weighs
evidence and presents multiple points of view
• An important dimension of argumentation in
letters in corpus: diagnostic and predictive
reasoning
– hearing loss was caused by mutation in gene (GJB2)
– if HD, then chance that others in family are affected
• Those parts of letter can be represented by
Bayesian (belief) network
Bayesian Network
History/proband
Age: child
History/mother family
history of deafness: no
Genotype/mother
one abnormal copy of
gene GJB2
Genotype/father 2
abnormal copies of
gene GJB2
Node Key:
Observable predispositional
Non-Observable
50%
Genotype/proband 2
abnormal copies of
gene GJB2
50%
Observable evidential
Genotype/sibling
2 abnormal copies of
gene GJB2
Links:
May increase
risk
Biochemistry/proband
Connexin 26: abnormal
Causal
Test/proband
Physiology/proband
normal chemical
equilibrium:no
Result/proband
GJB2 test positive
Symptom/proband
deafness
Finding/proband
facial defects: no
Symptom/sibling
deafness
Symptom/father
deafness
GenIE Project Summary
• Building demonstration system (GenIE) to
help genetic counselors write letters
• Using HCI to ensure effectiveness of general
argument presentation techniques
• Using AI to to model expert’s reasoning and
argumentation strategies
•Techniques will be applicable in many
domains to problem of computer-assisted or
automatic multimedia generation of effective
technical arguments for lay audience