Professional Documents
Culture Documents
T03 Screen
T03 Screen
ABSTRACT
The purpose of this tutorial is to introduce participants to the area of multimodal corpora,
covering central aspects of the way in which multimodal corpora are collected, annotated,
analysed and used to model human behaviour as well as to build multimodal systems. The
course will be a mixture of theoretical presentations, discussions of annotated video examples
and some practical annotation exercises, covering the following three parts:
1. Language and non-verbal behaviour in human communication
Theoretical approaches to multimodal communication from the literature will be presented and examples from video material will be shown and discussed. Methodologies for the
classification of different types of non-verbal behaviour will be introduced, and the interaction
of gesture with speech will be discussed.
2. Annotation models and procedures
Various approaches to gesture annotation will be described, covering issues related to
the identification, segmentation and interpretation of multimodal contributions. Examples of
multimodal annotation created with the ANVIL annotation tool (www.anvil-software.de) will
be shown. Reliability measures for multimodal annotation will be introduced.
3. Multimodal corpora, analyses and systems
Issues covered in this part include an overview of existing annotated multimodal corpora, a discussion of how multimodal corpora are analysed to gain insight into various aspects
of multimodal communication, how machine learning can be applied to multimodal data, and
how multimodal data are used to develop talking heads and embodied conversational agents.
!"
#$%&%'!$(!$)$*+',-!,'.+/)&!
/$!.&0$!1)00&$/1 /)$
02434
044526644786
934504864
72:8
;<;<
0123457897
3
0
9
0
!42"#
$!
0%
$
!
0123457897
3
0
7
0123457897
3
0
5
3 !"
""
#$#
"
%
!&
!'
0123457897
3
0
!" #$%&'()&*+,%$%-.
/
01
21
2
031
1
4&))5'%"+$%&'()&*+,%$%-.
10
06
217
0
/
4&)#5$!()&*+,%$%-(8%'#5$(+'*(
&5$#5$9.
:
6/1;
60
<=>?@ABBC
1
/1
1D
0123457897
3
0
#
0123457897
3
0
0
4
!"
#
$%
&
'(!)
'
%
0 !
42 !
"
#
!$
!
!
#
$
%
!
!
!
#
&!
!'(
!
!$
32)*+,-)
0123457897
3
0
98
0123457897
3
0
99
0123457897
3
0
97
! "#
$ % &
0123457897
3
0
9
3!"
!"#
$%&'&
!&''&!
$
$!('
)')*('&
&
$!!
0123457897
3
0
9
3
!
"#!
" #
"$!
0123457897
3
0
9
0123457897
3
0
9
C
A
EF2FFA
GABA
0123457897
3
0
7
0
!
"
#$%&'788()
''*!
+
,
''
-.
/
01
0123457897
3
0
7
0123457897
3
0
7
0123457897
3
0
7
0
0123457897
3
0
7
1
!"#
$$!%!#
3!
"%&'()*+,
$3-.%
"/01293495
6--%%%77!
-%! #89:;09";
</%=$9
0123457897
3
0
7
!"#""
!$%&
"
!'()
*+,-.+/012322451675175.,18955,2.:;1
<"!
=>
"?@@<<<))A
@<A"=BC
D74
E
0123457897
3
0
7
3
0123457897
3
0
9
0123457897
3
0
7
0123457897
3
0
12')*3)*4!00""!"#
156%7&')*3)*4"
"!-
0123457897
3
0
3
0123457897
3
0
0123457897
3
0
! "#
$%788&
! #
! #
! ''
0123457897
3
0
3!
0123457897
3
0
0123457897
3
0
!!
!
0123457897
3
0
8
#4$
%&'()(*+,-./
#0
%&%
0&&1&"'
2345634.
"
2)378+78 &9$&"
253(+. &&9$$&"
278,:78
243+(*;,-78
0123457897
3
0
9
!"##$$$%%
#$!&'()*+,
-$.
!"##$$$%%
#$!&'( 02
/01*823(
0123457897
3
0
7
%&'()*+(,-./0*-+12*-3/4526/7)8/
89)76(,-/
0123457897
3
0
!"
#
$
!$
$
%
!
"&
0123457897
3
0
%$&'9()7
*
+% 9,-,.
0123457897
3
0
"#
"&
'(
$
$'
)
'
)
0123457897
3
0
!2"#$%
&%%7889
'788
'788
5(
%
5$
)#%
*
$
0123457897
3
0
!!!#
$
!
!
&!'!!(
+,-./0123
0123457897
3
0
!
%
&#
'!'()
*+,-./.01/2
394567
8
9
:!
!
###$
$
#'
0123457897
3
0
8
ABCDE
;<=>?;@
9
:
@;FD;GG;@
D;HA<;@
9
0123457897
3
0
9
0123457897
3
0
7
0123457897
3
0
4
0123457897
3
0
0
0123457897
3
0
2
!""#$%&'$%(%)*+,$%-'%./0012%3456,7,8%9"!%:"%;5,584%<"==>,7?5:7",'%@,%ABCDEFGHBIHJDFKLFMNOPQHR/$%SS'%/RTRU/R10'%
!""#$%&'$%3'%<4665:"$%V'%&"W7,4,$%<'%X5Y5664::5%(%Z'%Z5887"%.[\\]2%^)4%;_;@X%<"#7,8%`?)4=4%a"6%:)4%,,":5:7",%"a%b44#c5?W$^>6,%
;5,584=4,:%5,#%`4d>4,?7,8'%@,%&'%<'%;56:7,%4:%5'%.4#*2%eCGMNLBfFGHgBDhBDFHIBDHeBfiGGNEKHjCLFEHeCGMNLBfFGHkilFmNBCD'%`S4?75%7**>4%
"a%:)4%@,:46,5:7",5%&">6,5%"a%35,8>584%n4*">6?4*%5,#%-Y5>5:7",'%`S67,846$%[\\]
o5:4*$%-'$%o4,78,7$%3'$%o64:)46:",$%@'$%<56,57",7$%3'%(%p":4665$%p'%./0]02'%qliHrLiDKiEOiHBIHstLuBGPvHgBKENMNBEHFEfHgBLLCENOFMNBEHNEH
wEIFEOt'%X4!%x"6Wy%?5#4=7?%Z64**'
-a6",$%z'%./0{/2'%|iPMCDiQH}FOiHFEfHgCGMCDi'%;">:",'
Z'%-W=5,%5,#%~'%b674*4,%./0]T2'%_,=5*W7,8%:)4%a5?4'%%8>7#4%:"%64?"8,77,8%4=":7",*%a6"=%a5?75%?>4*'%-,84!""#%<7aa*$%X4!%&46*4y%
Z64,:7?4U95$
V4,#",$%'%4*:>64%5,#%*S44?)y%:!"%5*S4?:*%"a%:)4%S6"?4**%"a%>::465,?4'%@,%;'n'%V4$%4#'$%BEmiDuFGHgBLLCENOFMNBEHFEfHFEKCFKi$%^)4%
958>4y%;">:",$%/01\$%SS'[\]U[[]'
V4,#",$%'%[\\{'%|iPMCDivHNPNuGiHOMNBEHFPHMMiDFEOi'%<5=c67#84y%<5=c67#84%_,7Y46*7:%Z64**'
V7SS$%;'%.[\\{2%|iPMCDiH|iEiDFMNBEHutHwLNMFMNBEHDBLHjCLFEHkilFmNBDMBHgBLhCMiDHglFDFOMiDHENLFMNBE'%o"?5%n5:",$%b"67#5$%
z7**46:5:7",'?"='
V7:5$%`'%.-#'2'%.[\\R2'%JBNEMNEKHliDiHGFEKCFKiQHOCGMCDiQHFEfHOBKENMNBEHLiiM'%97*#54%X&y%-6c5>='
;?X47$%z'%.[\\T2'%|iPMCDiHFEfHqlBCKlM'%<)7?58"y%_,7Y46*7:%"a%<)7?58"%Z64**'
;?X47$%z'%./00[2'%jFEfHFEfHLNEfvHlFMHKiPMCDiPHDimiFGHFuBCMHMlBCKlM'%<)7?58"y%_,7Y46*7:%"a%<)7?58"%Z64**'
;?X47$%z5Y7#%.?:"c46%/0]02'%qliHgBEOihMCFGHkFPNPHBIHFEKCFKi'%X4!%&46*4$%_`y%-6c5>=%%&"),%~74%(%`",*%@,?'
Z476?4$%<'%`'%./0R/UT12'%qliHgBGGiOMifHJFhiDPHBIHgHsHJiNDOi$%Y"*'%/U$%4#'%<)564*%956:*)"6,4%5,#%Z5>%~47**%Y"*'%]U1$%4#'%'%~'%
o>6W*$%<5=c67#84y%956Y56#'
Z"887$%@*5c45%.[\2%^)4%347?",%5,#%:)4%S)5c4:%"a%4*:>64$%54$%5,#%^">?)'%@,%wEMiGGNKiEMHNDMCFGHKiEMP34?:>64%X":4*%7,%<"=S>:46%
`?74,?4$%[\\/$%p">=4%[/0\[\\/$%[RTU[R'
n7=4$%o'$%5,#%`?)75:>65$%3'%./00/2'%4*:>64%5,#%*S44?)'%@,%n'`'%b4#=5,%5,#%o'%n7=4%.-#*'2$%CEfFLiEMFGPHBIHEBEmiDuFGHuilFmNBCD'%
956Y56#y%<5=c67#84%_,7Y46*7:%Z64**'
`?)4646$%V'%n'%%.[\\T2'%~)5:%564%4=":7",*%5,#%)"!%?5,%:)4%c4%=45*>64#%sBONFGHsONiEOiHwEIBDLFMNBE$%Y"'%{{$%,"'%{$%SS'%0T][0$%
z4?4=c46'
~>,#:$%~'%./0\T2%%|DCEfKifiDHhltPNBGBKNPOliEJPtOlBGBKNi'%347S78y%-,84=5,,'
0123457897
3
0
UNIVERSITY OF MALTA
Part 2
UNIVERSITY OF MALTA
Outline
Dias 2
UNIVERSITY OF MALTA
Dias 3
UNIVERSITY OF MALTA
Annotation tools
Dias 4
UNIVERSITY OF MALTA
An example: ANVIL
Dias 5
UNIVERSITY OF MALTA
Attributes and
values to be used in
the annotation
Dias 6
UNIVERSITY OF MALTA
Dias 7
UNIVERSITY OF MALTA
Dias 8
UNIVERSITY OF MALTA
ANVIL - windows
Dias 9
UNIVERSITY OF MALTA
ANVIL - elements
Dias 10
UNIVERSITY OF MALTA
Dias 11
UNIVERSITY OF MALTA
ANVIL - tracks
Dias 12
UNIVERSITY OF MALTA
ANVIL - groups
Dias 13
UNIVERSITY OF MALTA
UNIVERSITY OF MALTA
Dias 15
UNIVERSITY OF MALTA
Identification of gestures
Dias 16
UNIVERSITY OF MALTA
Identification, basic...
Dias 17
UNIVERSITY OF MALTA
Identification, basic...
Dias 18
UNIVERSITY OF MALTA
Segmentation
Dias 19
UNIVERSITY OF MALTA
Segmentation
Dias 20
UNIVERSITY OF MALTA
Segmentation
Dias 21
UNIVERSITY OF MALTA
Segmentation
Dias 22
UNIVERSITY OF MALTA
Segmentation, example
UNIVERSITY OF MALTA
Dias 24
UNIVERSITY OF MALTA
Dias 25
UNIVERSITY OF MALTA
MacNeill (1992)
Coling 2012, Mumbai
Dias 26
UNIVERSITY OF MALTA
Dias 27
UNIVERSITY OF MALTA
Dias 28
UNIVERSITY OF MALTA
Movement down-up
Jerk
HeadBackward
HeadForward
Tilt
SideTurn
Shake
Waggle
HeadOther
Dias 29
UNIVERSITY OF MALTA
Dias 30
UNIVERSITY OF MALTA
Gesture interpretation
Dias 31
UNIVERSITY OF MALTA
Gesture interpretation
Gesture
Meta.cup
Description
Speech sample
Form features
Meaning
Dias 32
UNIVERSITY OF MALTA
Gesture interpretation
Attribute
Values
FeedbackBasic
CPU, Other
FeedbackDirection
FeedbackAgreement
Agree, NonAgree
EmotionAttitude
Turn
Sequencing
Dias 33
UNIVERSITY OF MALTA
Gesture interpretation
Dias 34
UNIVERSITY OF MALTA
Gesture interpretation
Dias 35
UNIVERSITY OF MALTA
Gestural interpretation
Dias 36
UNIVERSITY OF MALTA
Gesture type
Corresponding speech
Emblem
Deictic
Iconics/
metaphorics
Beat
Stressed word
Dias 37
UNIVERSITY OF MALTA
Dias 38
UNIVERSITY OF MALTA
Reliability
Dias 39
UNIVERSITY OF MALTA
Reliability: inter-annotator
agreement
Dias 40
UNIVERSITY OF MALTA
Reliability: percentage
agreement
Dias 41
UNIVERSITY OF MALTA
Dias 42
UNIVERSITY OF MALTA
Dias 43
UNIVERSITY OF MALTA
P(A)-P(E)
K=
1 P(E)
Dias 44
UNIVERSITY OF MALTA
Jerks
Nods
20
Jerks
Coder A
Coder B
Nods
10
15
Dias 45
UNIVERSITY OF MALTA
1 P(E)
Jerks
Nods
K=
Coder A
P(A)-P(E)
20
Jerks
Coder B
Nods
10
15
Dias 46
UNIVERSITY OF MALTA
Dias 47
UNIVERSITY OF MALTA
Coder B
Nod
Jerk
Total
Nod
Jerk
Total
12
Dias 48
UNIVERSITY OF MALTA
Dias 49
UNIVERSITY OF MALTA
Conclusion
Dias 50
UNIVERSITY OF MALTA
References (1)
Dias 51
UNIVERSITY OF MALTA
References (2)
Dias 52
Part 3
Multimodal Corpora,
Analysis and Systems
Our goals
H
Coling 2012, Mumbai
H
Dias 2
This part
User
studies
Corpus
IPL
Machine
Learning
Dias 3
Beyond Corpus
Dias 4
Evaluation
Backchannels
Dias 5
MultiLis corpus
Dias 6
MultiLis corpus
Dias 7
Dias 8
Random strategy
Dias 9
Copy strategy
Dias 10
Dias 11
Gaze strategy
Dias 12
Dias 13
Dias 14
Corpus
Dias 15
Stimuli
Dias 16
Procedure
Dias 17
Results strategy
100
90
Average score
80
70
60
50
40
30
20
10
0
Copy
Random
Ward &
Tsukahara
Gaze
Dias 18
Results timing
Copy
Random
W&T
Gaze
P&P
PP&G
51
47
54
25
25
49
Precision (%)
12.77
9.26
4.00
40.00
22.49
Recall (%)
11.76
9.80
1.96
19.60
21.57
F-score
0.122
0.095
0.026
0.263
0.220
Number
Dias 19
Dias 20
Dias 21
Yucks
Dias 22
Dias 23
Dias 24
Dias 25
Schematic setup
Dias 26
Schematic setup
Dias 27
Dias 28
Dias 29
Dias 30
SWITCHING WIZARD OF OZ
Dias 31
Ingredients so far
Dias 32
Evaluation
Dias 33
Switching Wizard of Oz
Dias 34
Switching Wizard of Oz
Dias 35
Switching Wizard of Oz
Dias 36
Schematic setup
Dias 37
Experiment: results
Dias 38
Experiment: results
Dias 39
Dias 40
Dias 41