Speaking patterns MAS.662J, Fall 2004

1
Speaking patterns-MAS.662J, Fall 2004

• Diane Hirsh Xian Du
• Dec-07-2004

2
Outline

• Introduction
• Data at hand
• Objectives
• Applied Methods
• Results
• Conclusion and Comments
• References

3
Introduction

• Group debate always leads to only two kinds of
  final decision right or wrong
• Debating Member always influences each other by
  different speaking patterns
• The speakers pattern and other speakers
  influences lead to the final decision of the
  debate
• Identifying those patterns and influences can be
  helpful to the prediction of the debate result
  and members option

4
Data at hand

• Raw data is Speaker ID and stamp time
• Labeled data indicates the initial position and
  final outcome of the speakers

Fig. 1 talking sequences of the four members in
group study_07_task1
5
Objectives

• Find the distinct feature to discriminate the
  right-decision from wrong-decision group
• Predict the winner of the project
• Tell the individuals position

6
Applied Methods

• Preprocessing
• Turn 1,2 for each unit of time we
  estimate how much time each of the participants
  speaks, the participants who has the highest
  fraction of speaking time is considered to hold
  the turn for that time unit. For a given
  interaction, we can easily estimate how a pair
  participating in the conversation transitions
  between turns.

7
Applied Methods

• Recognition techniques
• Parzen Window Linear Discriminant Function with
  one-leave-out validation
• Hidden Markov Models (HMM)

8
Parzen Window Linear Discriminant Function with
one-leave-out validation

• Goal
• - to discriminate the right-decision group
  from wrong-decision group
• Assumption
• - Wrong decision group has a wrong density
  function (Parzen window)
• - There is a hyperplane H to divide the
  turns space into half spaces right or wrong
• Group 07 and 12 are two wrong groups in the
  training groups

9
Results for Parzen Window Linear Discriminant
Function application

• Parzen window
• - 07 group 6 in 10 right groups are
  identified but Minimum error rate0.5
• - 12 group fail (5 in 10 and 0.5)
• Linear Discriminant
• 07 group 8 in 10 right groups are identified
  with Minimum error rates 0.0570.47, AVG0.269
• - 12 group fail (5 in 10 and 0.5)

10
Hidden Markov Models (HMM)

• Single HMM
• Identify Group option (wrong/right)
• Parallel HMM
• Identify Group option (wrong/right)
• Identify members state option (probability of
  the final decision)
• Influence model
• - Improve the result of parallel HMM by
  considering the influence between members in the
  group

11
Implementation of HMM

• Assumptions
• - each member in one group has influence on
  others by turns amount and more turns
  contribute to higher influence.
• - each member retains its initial state or
  changes to be opposite. The transition is
  strictly one direction.
• Initializations
• - randomize the initialization of transition
  matrix while keeping the HMM strictly
  left-to-right.
• - two states for each group right or wrong
  two observation symbols 0 or 1.
• - The initial states of each member in the
  group are set according to the initial position
  in labeled data (e.g.1/4)

12
Implementation of HMM
13
Result for HMM

• Single HMM for right-wrong groups separation
  cannot find wrong groups
• e.g..
• training data 1,2,3,4,5,6,8,9,10,11 and
  7,12
• testing data 1,2,3,5,6,7,8,9,10,12 and
  4,11
• Confusion Matrix for the Test Data
  (test 9)
• Recognized as right Recognized
  as wrong
• Class 1 8
  0
• Class 2 2
  0

14
Result for HMM

• Parallel HMM (group 8 and group 11 used)
• Recognition accuracy 54.4
• Members state transition matrix
• - group 08 1/4 meet labeled data
• - group 11 3/4 meet labeled data
• Influence HMM model
• Recognition accuracy 61.0
• Members state transition matrix
• - group 08 2/4 meet labeled data
• - group 11 3/4 meet labeled data

15
Conclusion and Comments

• The data set is not friendly for HMM because
  different training group has different members
  and each group has only one speech sequence
• Influence model improves the HMM recognition
  accuracy but its random initial state probability
  limits its application in this project (it needs
  more training) and its result up to now failed to
  find the winner
• Linear discriminant function recognizes some
  right-wrong group well but not all (more data
  needed for the testing)
• The length of talking time varies a lot among
  different groups which limits the recognition
• More features may be helpful for this project.

16
References

• Tanzeem Khalid Choudhury, Sensing and modeling
  human networks, PhD thesis, MIT, Cambridge, MA,
  2004
• Chalee Asavathiratham, the influence model a
  tractable representation for the dynamics of
  networked Markov chains, PhD thesis, MIT,
  Cambridge, MA, 2001
• A Pentland, Learning communications-understanding
  information flow in human networks, BT technology
  Journal, vol. 22, No4, October 2004
• Shi Zhong and Joydeep Ghosh, A new formulation of
  coupled hidden markov models, A new formulation
  of coupled hidden markov models, Tech. Report,
  June, 2001
• YongHong Tian, etc. Incremental learning for
  interaction dynamics with the influence model,
  IEEE, www-2.cs.cmu.edu/dunja/LinkKDD2003/papers/T
  ian.pdf
• Lawrence Rabiner and Biing-Hwang Juang,
  Fundamentals of speech recognition, Prentice
  Hall, 1993