Dominance detection in Meetings Using Support Vector Machines

1
Dominance detection in Meetings Using Support
Vector Machines

• Rutger Rienks and Dirk Heylen
• University of Twente
• MLMI 2005

2
Outline

• Dominance
• Dominance Judgements
• Towards automatic dominance detection
• Results
• Conclusion

3
Cooperative

• For a meeting to be effective, participants
  should be cooperative
• The Grices cooperation principle is founded on
  four maxims Grice,1975
• Speak no more or less than required
• Only say things for which you have evidence
• Only say things relevant for the discussion
• Formulate such, that you are clearly understood

4
Uncooperative

• The chairman of a meeting should e.g.
• Make sure that everybody can get the floor
• Cut off people when speaking too long
• Intervene when contributions are irrelevant
• People that are too dominant frustrate the
  cooperative principles and hence the meeting
  process
• Can we automatically detect if one person in a
  meeting is more or less dominant than someone
  else?

5
Dominance

• The Dominator is a group member trying to
  assert authority by manipulating the group or
  certain individuals in the group. Dominators may
  use flattery or proclaim their superior status to
  gain attention and interrupt contributions of
  others Hoffman,1979

6
Dominance

• Behavioral features displayed by people that
  behave dominantly Bales Cohen,1979, Bales,
  1951

7

• But, do people agree on
• the concept of dominance?

8
Dominance Judgements

• Corpus of eight four-person meetings (95 minutes)
• Ten people to rank the meeting participants of
  four meetings on their conveyed dominance
• Result a total of five rankings for every
  meeting

9
Dominance Judgements
µ Reranking of the sum of all annotators for
each participant s The sum of the differences
for all annotators with µ for each participant
10
Dominance Judgements

• When comparing the variance of the judges with
  the variance resulting from randomly generated
  rankings, the distribution of the variance of the
  annotators significantly differs (plt0.001) from a
  distribution with randomly generated rankings.
• So, people seem to agree.

11
Towards Automatic Dominance Detection

• Dominance can be regarded as a higher level
  concept that may be deduced automatically from a
  subset of lower level observations Reidsma et
  al., 2004.
• Most of the features weve seen from Bales and
  Bales and Choen are hard to operationalize and
  measure. (c.f. Alienated and Purposeful)
• For our classifier we considered some common
  sense features that possibly might reveal us
  something about the dominance of a person in in
  meetings.

12
Towards Automatic Dominance Detection

• Types of extracted features from the meetings
• The speaking time in seconds
• The number of words spoken in the whole meeting
• The number of turns in a meeting
• The number of times addressed
• The number of times privately addressed
• The number of successful interruptions
• The number of times interrupted
• The ratio of successful interruptions and the
  number of times being interrupted
• The number of times the floor is grabbed by a
  participant
• The number of questions asked

13
Towards Automatic Dominance Detection

• The obtained feature values were normalized and
  made comparable by mapping them on a (High,
  Normal, Low) scale
• The average judgement rankings were mapped onto
  the same (High, Normal, Low) scale and used as
  class labels
• This resulted in a data-set of 32 samples with
  twelve samples receiving the class label High,
  ten Normal and ten Low.
• We define our baseline performance as the share
  of the most frequent class label (High), which
  was 37.5 of all labels.

14
Towards Automatic Dominance Detection

• We aimed to predict the dominance level of the
  meeting participants with the least number of
  possible features. As obtaining some features is
  expected to be easier than obtaining all
  features.

15
Towards Automatic Dominance Detection

• To decrease our amount of features we used Wekas
  SVM Attribute Evaluator
• The top five features appeared to be
• The number of floorgrabs
• The number of turns
• The number of successful interruptions
• The number of words used
• The number of questions asked

16
Results

• The classifier obtained the highest performance
  (75) using the best two features (10 f.c.)
• The Confusion Matrics
• The 90 confidence interval for our classifier
  lies between a performance of 62 and 88. Having
  a lower bound higher than the 37.5 baseline.

Actual
Predicted
17
Conclusion

• Aware of the fact that our sample size is
  relatively small and that not all meetings follow
  the same format, we do think that our results
  suggest that
• It is possible to have a system analyzing the
  level of dominance of the meeting participants.

18
Further research

• Run tests on more data
• Use semantically oriented features
• Try to apply a system in real time
• Investigate te impact of the results on various
  types of players (chairman, participants, agents)