Poster

1
Humor Prosody Analysis and Automatic Recognitio
n For F.R.I.E.N.D.S Amruta Purandare and Diane
Litman University of Pittsburgh EMNLP 2006, S
ydney, Australia
1. Motivation Need social intelligence in comput
ers Approaches Affect, Personality, Humor? St
ate of the art in Computational Humor
Humorous Text (Acronyms, One-liners, Wordplays)
Lexical cues (Alliteration, Slang, Antonymy)
Our contribution Humor Detection in Spoken Con
versations Do Prosodic cues (e.g. Pitch, Intensi
ty, Tempo) help?
6. Humor-Prosody Analysis (Results)
Most prosodic features show significant (p
for a t-test) differences between Humor and Non
-Humor groups Humorous turns show higher Max,
Range, Std-Dev in Pitch and Energy, higher Temp
o and smaller Internal Silence
7. Gender effect on Humor-Prosody (Results)
Accounting for gender differences with 2-way
ANOVA The test shows Humor effect on prosod
y adjusted for Gender Gender effect on prosody a
djusted for Humor Interaction effect between Gen
der and Humor i.e. if the prosodic style of ex
pressing humor is different for Males and Fem
ales Findings Significan
t effect of Humor even when adjusted for Gende
r 2) Significant effect of Gender, but only P
itch features show the Interaction Effect. i.e
. males and females use different Pitch variatio
ns while expressing Humor
2. FRIENDS Corpus 75 Dialogs from a classic TV-c
omedy FRIENDS 2hrs of Audio Text transcripts
from http//www.friendscafe.org/scripts
Humorous turns are followed by laughs
Automatic labeling using laughs
Corpus size 1629 turns 714 Humorous, 915 Non-
Humorous 6 Main Actors (3 Male, 3 Female), 26 Gu
est Actors
Y significant effect N non-significant ef
fect
3. Example Dialog Rachel Guess what? no Ros
s You got a job? no Rachel Are you kidding?
I am trained for nothing! yes
Rachel I was laughed out of 12 intervi
ews today. no Chandler and yet you are surpr
isingly upbeat! no Rachel Well, you would be
too, if you found John Davids boots on sale,
50 off... yes Chandler Oh how well
you know me! yes Rachel They are m
y new, I dont need a job, I dont need my
parents, I got great Boots, Boots! yes
yes Humorous Turns no
Non-Humorous Turns
8. Humor Recognition (Results)
Supervised 2-way classification
Results above baseline (56.2)
Results consistent for genders
Marginal improvement higher for males
Decision tree shows that the algorithm picked
mostly prosodic and speaker features in the fi
rst 10 iterations
4. Features Borrowed from emotional speech liter
ature Prosodic (13) Pitch (F0) Mean, Max, Ran
ge, Std-Dev Energy (RMS) Mean, Max, Range, Std-
Dev Temporal Duration, Internal Silence, Tempo
Lexical (2011) all Words Turn Length
(words in the turn) Speaker ID (1)
9. Conclusions Future Work Humor recognition
in spoken conversations Data Dialogs from a
classic comedy TV show, FRIENDS
Used laughs for automatically labeling humorous
turns Humor-Prosody Analysis Humorous turns
show higher peaks and variations in pitch and
energy, and higher tempo, compared to
non-humorous turns Gender Effect Most featur
es show humor effect even when adjusted for
gender Only pitch features show the interaction
effect Results Promising, 8 over the baseline
with all features Humor detection easier for ma
le speakers than for females Future Pragmati
c features e.g. Ambiguity, Incongruity, Expect
ation-Violation etc.
5. Feature Extraction using Wavesurfer