Lexical%20markers%20of%20conversational%20disagreement%20in%20psychotherapy%20interviews

1
Lexical markers of conversational disagreement in
psychotherapy interviews

• Shuki Cohen
• New York University

2
Why Modeling Disagreement? Linguistic Features

• High Frequency of feedback in conversation good
  time resolution
• High Frequency of marker tokens in spoken
  language small sample size, high reliability
• Relatively small, ritualized, formulaic lexicon
  token-based statistics

3
Why Modeling Disagreement? Psychological Features

• Therapeutic Alliance best predictor of therapy
  outcome
• Increasing Patient Responsibilities
  Client-Centered managed care.
• Indeterminancy of the impact of Psychoanalytic
  interpretations when is the therapist right?

4
Experimental Design

• 39 hours of recorded psychotherapy
• 15 patient-therapist dyads
• 3 successive sessions per patient
• 2384 therapist utterances subsequent patient
  response
• training corpus 1196 therapist-patient exchanges
  (shuffled)
• Validation corpus 1188 therapist-patient
  exchanges (shuffled)
• 4 different raters for each corpus. All native
  speakers of American English

5
Results

• Average Inter-rater reliability ?0.76
• Correlation of model with human ratings 0.85

6
Markers of agreement
Word Beta Std. Err. Std. beta Signif.
Yeah 0.946 0.033 0.496 Plt0.01
Mm-hm 1.213 0.068 0.318 Plt0.01
Ok 1.131 0.068 0.295 Plt0.01
Right 0.931 0.063 0.254 Plt0.01
Yes 1.288 0.193 0.111 Plt0.01
Fine 0.862 0.164 0.088 Plt0.01
Exactly 1.089 0.249 0.073 Plt0.01
Absolutely 0.558 0.272 0.034 Plt0.05
Good 0.309 0.195 0.026 P0.113
7
Markers of disagreements (I)
Word Beta Std. Err. Std. beta Signif.
No -0.957 0.056 -0.286 Plt0.01
Um -0.255 0.050 -0.088 Plt0.01
A -0.299 0.071 -0.072 Plt0.01
Not -0.296 0.079 -0.069 Plt0.01
Actually -0.543 0.160 -0.057 Plt0.01
To -0.248 0.075 -0.056 Plt0.01
Really -0.303 0.102 -0.051 Plt0.01
Because -0.258 0.091 -0.048 Plt0.01
Like -0.150 0.055 -0.046 Plt0.01
Uh -0.214 0.081 -0.045 Plt0.01
Dont -0.183 0.073 -0.049 Plt0.05
You -0.210 0.084 -0.057 Plt0.05
Was -0.131 0.059 -0.042 Plt0.05
8
Markers of disagreements (II)
Word Beta Std. Err. Std. beta Signif.
That -0.121 0.055 -0.038 Plt0.05
Well -0.139 0.065 -0.037 Plt0.05
Honestly -0.736 0.363 -0.035 Plt0.05
Hm -0.373 0.187 -0.034 Plt0.05
And -0.099 0.052 -0.032 Plt0.1
Mean -0.152 0.086 -0.032 Plt0.1
Definitely -0.505 0.305 -0.028 Plt0.1
9
Agreement Preference
10
Phasic nature of conversational
agreements-autocorrelation structure
11
Towards validation of regression-based model

• Disagreement lexicon bigger and more diverse than
  agreement lexicon
• Coefficients magnitude (mostly) in accord with
  common intuition of strength of agreement or
  disagreement
• Preference for agreement vs. disagreement
• Phasic nature of agreement episodes.

12
Instances of compromised performance

• Agreement with negative statements
• T so she wasnt, she wasnt living there?
• P no she was on vacation like all the people
  out there are
• No feedback (continuation, rephrasing, stressing
  etc.)
• repetition, stuttering
• P yeah, yeah, yeah
• Incongruent response
• T so it must have been painful
• P no, it was painful.

13
Limitations

• Psychotherapy conversations
• Low familiarity - early stage of therapy
• Semantic level, prosody/non-verbal information
  missing
• larger context/background for conversation is
  missing
• Detection of common ground buildup questionable
• Gender difference in both appraisal as well as
  performance of agreement (raters and speakers
  alike).
• Agreement? Acknowledgment? Discourse markers?

14
Conclusions

• Computer able to unpack a significant portion of
  human intuitive judgment regarding agreement and
  disagreement
• Computerized scale for modeling agreement has
  high reliability and stability, and sufficient
  validity for certain studies
• Analysis utilizing the computerized scale able to
  uncover repetitive patterns in conversations that
  are stable across topics and context and are
  (presumably) unconscious.