Slide deck used to present and defend my master's thesis project. The project is detailed in a paper published in the Proceedings of the 2009 Conference on Intelligent User Interfaces (http://doi.acm.org/10.1145/1502650.1502696).

1. A Multimedia Interface for Facilitating
Comparisons of Opinions
Lucas Rizoli
Supervisor Giuseppe Carenini
University of British Columbia
February 11, 2009

2. Opinion data is
abundant and useful,
but analysis is expensive and difficult
2

3. Our interface
supports analysis of opinion data,
particularly comparison across entities
3

4. It supports analysis by
visualizing the data and
summarizing notable comparisons
4

5. Our user study shows
the visualization is usable,
the summarizer’s choices are human-like
5

6. 6

7. Text mining
(Carenini, Ng, & E. Zwart, 2006)
7

8. 8

9. 9

10. 10

11. 11

12. 12

13. +2 +1
+2 –3
13

14. Spore
Gameplay Technology DRM
Creature creation Graphics
Cell phase Sound
Civilization phase System requirements
14

15. vs.
15

16. 16

17. Applications
● Consumers
● Intelligence
– Competitive analysis
– Forecasts
● Research
– Survey, questionnaire
17

18. 18

19. 19

20. Controversiality
(Carenini & Cheung, 2008)
20

21. 21

22. 22

23. 23

24. (Dis)similarity of comparisons
● Adapt stats to pairs of distributions
● Aspects of a comparison
– counts
– means
– contros
– dists
24

25. 25

26. 26

27. 27

28. 28

29. 29

30. 30

31. 31

32. 32

33. 33

34. 34

35. Interpreting aspects
● Aspects in range 0–1
● Interpretation
– means 0.6 != contros 0.6 != dists 0.6
35

36. 36

37. 37

38. Selecting comparisons
1. Filter out low-count comparisons
2. Rank by # of extreme aspects (VD, VS)
3. Realize top comparisons
• Sentence templates
38

39. 39

40. Study goals
● Evaluate content selection strategy
– Matches human selections?
● Better than baseline?
– Humans like selections?
● Usability of visualization
40

41. 41

42. 42

43. Baseline selection strategies
● Naïve
– Randomly select which and how many
● Semi-informed
– Likely select same how many as subjects
– Randomly select which
43

44. Matches human selections?
44

45. Humans like selections?
0.500 0.449 0.454
45

46. Humans like selections?
0.500 0.449 0.454
● Roughly 20% difference
● 60% of system selections marked as “good”
46

47. Usability
47

48. Opinion data is
abundant and useful,
but analysis is expensive and difficult
Our interface
supports analysis of opinion data,
particularly comparison across entities, by
visualizing the data and
summarizing notable comparisons
Our user study shows
the visualization is usable,
the summarizer’s choices are human-like
48

49. Thank you
49

50. Future work
● Tune thresholds and aspects
● Analysis of human reasoning
● Machine learning for selection
50

51. Future work
● More evaluation of visualization
– Interaction
– Deeper heirarchies
– Different data
– Insight
● Multiple entities
● Improved summarization
– Visual cues
51

52. Future work
● Machine learning–based selection
– Trained on study data
– Which
● Regression on comparison selection scores
– How many
● Max # of comparisons (2)
52

53. Sentence construction
● Always main claim: counts
● All other aspects relate to counts
– Support: same (dis)similarity as counts
– Contrast: different
● Always mention means
● Mention contros, dists when they are extreme
53

54. counts
rt
co
o
pp
nt
u
ra
s
st
means contros dists
54

55. counts
rt
co
o
pp
nt
u
ra
s
st
means contros dists
55

56. counts
rt
co
o
pp
nt
u
ra
s
st
contros means
56

57. Study
● 36 subjects
– 24 female
– 19–43 years old
● 22 different pairs of entities
– Subjects saw ~4 each
57

58. Data generation
● No existing dataset
● Generated similar to existing datasets
– Distribution, modality
● Explore space of possible data
– Too large
– Representative of larger space
58

59. Generated data
● Generic camera features
– Consistent with scenario
● Simple heirarchy
– Reduce visual and task complication
59

60. Generated data
● 9 comparison types
– Constraints on aspects
– Range of support/contrast
● 22 summary cases
– Summaries by type
– Range of
● overall,
● how many,
● which,
● others.
60

61. 61

62. 62

63. 63

64. 64

65. 65

66. 66

67. 67

68. 68

69. 69

70. 70

71. 71

72. 72

73. 73

74. 74