Talking probabilities: communicating probabilistic information with words and numbers

1
Talking probabilities communicating
probabilistic information with words and numbers

• Silja Renooij, Cilia Witteman
• Presented by Jiangbo Dang

2
Outline

• Motivation
• Related works
• Four experiments
• Results and discussions
• Conclusion
• Additional references

3
Motivation

• Numerical probabilities is often considered a
  major obstacle in the construction of a Bayesian
  Network
• Verbal probabilities are perceived as more
  natural than numerical probabilities, easier to
  understand and communicate and better suited to
  convey the vagueness of opinions
• A combination of a numerical probability scale
  with a verbal probability scale may help to
  remove the mentioned obstacle

4
Why obstacle

• On the input side, it may be a prohibitive
  quantity of probabilities for all variables in
  Bayesian network. Experts are reluctant to
  express them in numerical form, they find it
  difficult to attach a number to their beliefs.
• On the output side, explanation of the results in
  term of variables with numerical probabilities
  may be un- comfortable.

5
Related work

• Translate numerical probabilities into words and
  vice versa
• Compare Verbal probability expressions with
  numerical probability expressions under certain
  situations (Physician prefer verbal to number)
• Limit the number of verbal expression in a scale

6
Four experiments

• Collect a list of commonly used verbal
  probability expressions
• Determine a rank order for probability
  expressions from previous experiment
• Rate difference between those expressions and
  develop a scale
• Test the translations of the verbal probability
  expressions into the calculated numerical
  probabilities.

7
1. Collect verbal probability expressions
8
2. Rank probability expressions
9
3. Rate distances between those expressions

• Pairwise comparison (28 pairs for 8 expressions)
• Produce co-ordinates for expressions, those
  co-ordinates best fit the distances given by
  different subjects
• Map co-ordinates onto a probability scale by
  setting certain and impossible representing 100
  and 0

10
3. Rate distances between those expressions
11
3. Rate distances between those expressions
12
4. Test probability scale
13
4. Test probability scale

• Compare the decisions subjects made when they are
  using probability in verbal form to their
  decisions when they are using numerical
  probability.

14
Conclusion

• A probability scale with mutually exchangeable
  verbal and numerical probability expressions is
  helpful in BN constructions
• Numbers may be just as vague as words for
  expressing uncertainty
• Shortcoming Dutch subjects and Dutch words

15
References

• Cilia Witteman, Silja Renooij. Evaluation of a
  verbal-numerical probability scale
• L. C. van at el. Probabilities for a
  probabilistic network a case study in
  oesophageal cancer