Content uploaded by Sungbok Lee
Author content
All content in this area was uploaded by Sungbok Lee
Content may be subject to copyright.
An Analysis of Multimodal Cues of Interruption in Dyadic Spoken Interactions
Chi-Chun Lee, Sungbok Lee, Shrikanth S. Narayanan
!""#!
{chiclee,sungbokl,shri}@usc.edu
Abstract
$ %
&$
' ( & change
activeness, $
' &
&
$ &
)*+
&&
&
$
1. Introduction
,
$ '-.
/*0 &
.. &
. . $ 1
&
-
-$
2 , & 3
/)0$ 4 .
($ &
$
&&
-. 3
$ 5 /60
& 7
$
& &
.$ 4
& &
.&
&$ '
$ 2, 8/90 (
, &
$ :& ,
$ 2
.$8;/<0(
&&
. $ 4
$
;
&
/= >0$ ; &
$'?4@/#0&
- $ : (
( &
$ 2
&
$
& &
$
' (&
) ,
6 &.
9$
2. Research Methodology
4 & &
7 $ :
A: /!0
B
& &
C$ '
D D
$ :
&
& /*"0 3
& & &
$
2.1. Database and Annotation
: ?4@ $&
,$ ' &
-
$ :
$ '
$ =*
.) <6 6
& D
x , y , z
. $ ' . &
&
$ '
D & $ ' &
$
6$*$6 &
$
&
$ ' &
D$ 4
& ,
$ ' E'/**0&
&
$
&?4@**66
*=&D.$
&
. & & .
F
. $ BC
/*)0$'
( /*60$,
&
Competitive Interruption
?F)GH G$
2FH&$$$
Cooperative Interruption
?F&& & H&$
2FH $$$
& &
$ %. ,
$'*
$ '
&.$
'*FSummary of Interruptions
'
? 6!*= *>> <=)6
2 ==6) 6**# !><"
' *"<9# 9#)< *<6>6
2.2. Feature Extraction
' , F
F F
$%
&
&
$ :
$ ;& &
.G
.$ ' ,
$
&
change activeness,
& &
$ Change
&
& $ Activeness
&$
2.2.1. Speech-Intensity
1& *" &&&
6" @ /*90$ '
•?,D-&
•?,&
•?D-&
•?&
•change D-&
•activeness &
changeD-&
Iw=Ifew−Ifsw
*
&
few, fsw
&$activeness
)
Iov =
∑
i=fsovI
feov
∣
Ii−Ii−I
∣
Teov −Tsov
)
&
feov , fsov
Teov , Tsov
$
2.2.2. Hand Motions
; .
?4@$'
•1activeness D-&
•1change D-&
•activeness D-&
•change D-&
6 .
.
.$'&activeness
Vk=
∑
i=fswI
few
xi−xi−IJyi−yi−IJ zi−zi−IJ
Te−Ts
6
&
xi, yi, z i
.
Te−Ts
&$'
&
&
$
;&
-
&. &
Vk
Vk
$ ' &
&
$Activeness,
rw
rw=V1
〈V〉,
&
〈V〉 =
∑
k=1
totw
Vk
totw
9
V1
&
6
totw
&$
Change &
&
&$ ' 3 &
$
'
$ ; & .- & . K 9
x , y , z
.
2*$
2 * D
x , y , z
x , z
, &$ G
$'
2*Fxample of clustering of right hand
motions into 4 regions
,
$ & &
&
$ ' change
F *
& & " &$ '
&
$
2.2.3. Disfluencies
:. *
" &
$:
&
&
$ ' &
(/*<0
•false start: .
•repetition: .
•filled pause: BC BC BC BC
2.2.4. Feature Normalization
( &
D &
.-$ 2
F
. &
Fref
Fref =
∑
sbj
∑
neu
F
numsbj∗numneu
<
&
sbj
neu
D .
$
numsbj
numneu
D D$
'(
Fnorm
Fnorm =F
Csbj
&
Csbj =
∑
neu
Fsbj
numneu
∗1
Fref
=
&
Csbj
(D$'
(
&($
3. Results and Discussion
' , & &
3
•Does each feature listed in Section 2 behave differently
for the two types of interruptions?
•Can we obtain a better discriminating power by
incorporating multimodal cues?
; two samplet-test, two proportions test,
fisher's exact test- (,
3$
3.1. Hypothesis Testing
4
3 &
$
. &
- .
$:,
$ ' )
)$
3.1.1. Speech-Intensity
'
D- &
& &
$ :
$
,
$ ' &
( $ ' change
activeness &
&.'&
$
3.1.2. Hand Motions
' ' )
$ '
$ E change fisher's exact
test &two sample
proportions test$ : &D .
L
.&$ % &
&&&$
3.1.3. Disfluencies
'p-value two proportions test$')
&
$
:(
&
.
&
.3.$
3.2. Discriminant Analysis
&
@&
•Intensity-only Features
•Hand Motions-only Features
•Combination of Both Modalities
' 6
$ '
&
6$*$ 2 ' *
& .
'): Summary Results of Interrupting Utterances
M ;?
?, ?
E
Word Overlap Word Overlap Left Right Left Right
=>$>< >"$# ="$9# =*$9! !*$!) **$<> )$#= )$<< *< *< 69
=<$9) =#$66 <!$6# ="$)* >*$6! #$"" *$=< )$*< * 6 >
p-value 0.05 0.013 "$*# "$"! 0.007 0.02 0.02 "$)< 0.006 "$"> 0.016
&
$
$ ' %,G ?
"$"<
$'
&$
3.2.1. Intensity-only
' ,
change activeness. ':.G
"$!"9 G
"$">"$'
& ' 6 #!$=+
)#+
$ :
($
3.2.2. Hand Motions-only
' G activeness change
G change. G changep-
valueN"$"< &
$' :.G "$#<!
G "$"*9$2 '
6
$
,
$
3.2.3. Combination
' :.G "$>#)
"$*"$2'6
$
;&
*9+ )*+
>!$< +$ '
(
$-&&
F,
Gchange. % D & &
>*$)+ &
$
4. Conclusions and Future Work
4
$ ' &
&
$ :
$$ & &
$
$' &
&
$4
6 &
$ ' (
&
$
&&.&
?4@
$'&&
& &
&.$ &
$;&
/*=0 &
,
& $ 2
&.
& .G
$ '
&
$& &
$'
&.$
5. Acknowledgment
' & E2
$
6. References
/*0 $ B1'..'
C Journal of Personality and Social
Psychology,)6F)#6-!) *!>)
/)0 $ % @$5$ 5 $ E O1-
-
OICASSP ; ; )""> $
) $=#<-=##$
/60 P$ 5$ B F
&- -
C Journal of Pragmatics *9 ##6-!"6 *!!"
/90 $ 8 BM( . F
C Proc. Second SIGdial
Workshop on Discourse and Dialog $*= *-*" )""*
/<0 2$8$@$;$ B1
C Proc. NAACL HLT 2007 1E8
)""> *>-)9
/=0 $ ?E Hand and Minds: What Gestures Reveal about
Thoughts,$@ *!!)
/>0 $; B$;?
C AISB ; Q$ )""<$
/#0 $% ?$% $$ $Q(( $?& $Q
P$E$ $ $$E O?4@F
O P
1 )"">$
/!0 $; A $ : B, 1
C Language and Sex: Difference and
Dominance,%' E; 1&
?FE&; *!><
/*"0 5 4. 1 - B?
F , ?
C Social Psychology Quarterly, $ =< $*
6#-<<
/**0 :$? 2 Syllabification Software, ' . E
@5 E
' P*!!>$FLL&&&$$LLL
/*)0 ?$ Q BR
C Proc. of the Eurpspeech,$*6=>R*6>" )""*
/*60 ;$A$ 8-8 1$8 $ 8$? 8$A
B F
'SC Journal of Intercultural
Communication Research,$69 $9 )66-)<9 $)""<
/*90 @$%$:. B@
C @
E ' 1 *6) *!!=
FLL&&&$$
/*<0 @$$;P$2$$B
.F?.T .
C Computational Linguistics )<9F<)<R<>* *!!!
/*=0 $% A$ ?$5 $ E $E
O1,F
O '
@ $*< $6 $*"><-*"#= ?)"">$
Table 3: Summary of Classification Result
4
*""$"+ "$"+ =<$>+
- #!$=+ )#$"+ =#$<+
;?- <9$)+ ##$"+ =<$#+
#!$=+ ="$"+ 79.5%