3. Scenario
As a group were discussing a new design problem.
Charlie suggests using red circles.
... Lucy shoots it down, says “That’s stupid” ...
5 minutes before the end of the meeting, Lucy says,
“I’ve got it! I’m a genius ... red circles!”
4. Scenario
As a group were discussing a new design problem.
Charlie suggests using red circles.
... Lucy shoots it down, says “That’s stupid” ...
5 minutes before the end of the meeting, Lucy says,
“I’ve got it! I’m a genius ... red circles!”
5. Scenario
As a group were discussing a new design problem.
Charlie suggests using red circles.
... Lucy shoots it down, says “That’s stupid” ...
5 minutes before the end of the meeting, Lucy says,
“I’ve got it! I’m a genius ... red circles!”
6. Scenario
As a group were discussing a new design problem.
Charlie suggests using red circles.
... Lucy shoots it down, says “That’s stupid” ...
5 minutes before the end of the meeting, Lucy says,
“I’ve got it! I’m a genius ... red circles!”
7. Speech recognition is not perfect
20-30% word error rate in normal conditions
[Munteanu 2006]
Natural language processing is hard
[Rosenfeld 2000]
8. Speech recognition is not perfect
20-30% word error rate in normal conditions
[Munteanu 2006]
Natural language processing is hard
[Rosenfeld 2000]
9. Human-Computer Dialog
Techniques to identify salient moments in unstructured transcripts
by leveraging human knowledge with the computational affordances
of computers
er o i o n
s
h
ta
u
re
co du s
tto
at v
st
utahs canyons colorados park mountain lake sandstone
ne
nw s
ri
ra
oo ct i
wi
nc
s
d on
h
ld
n ch
l
er
ai
w i l b i t a t s o sio n
ra
n ev ada
tr
ne
d er n ess
canyons trail mountain bikes wilderness hikers
ss
er
p ar
er o s
ut a
nevada
k
n
io
h
ion
os
rid
ha
at v s
er
er
ne
s
m ot
s
deputy rights motorcyclists atvs
n es
v ad
u t ah
p ar k
orb
r
il
d es
a
ut ca tr a
w ilr a
ah ny
ik e
i ls
t
t
on
r es
ar l an
nevada
ch
fo
ds
es r e
bik es
cr
va ad a
p ar k ea
park
montana bitterroot
ti o
a
n edv
n ne
al
ne
cr o
e r o s i on
13. Forms of Input
Speech Transcription Software
Participant Tactile Feedback
Wikipedia / Explicit Semantic Analysis (ESA)
[Gabriovich 06]
14. Generating Clusters with
Explicit Semantic Analysis (ESA)
Query: ‘forest preserves in Utah’
1 U.S. National Monument (732.787)
2 Utah Lake (646.047)
3 United States Forest Service (584.821)
4 Price, Utah (575.731)
5 Red Deer (469.844)
6 Colorado (453.202)
7 Protected areas of the United States (452.932)
8 Utah (451.928)
9 Western United States (431.971)
10 Utah County, Utah (427.949)
15. Generating Clusters with
Explicit Semantic Analysis (ESA)
Query: ‘forest preserves in Utah’
1 U.S. National Monument (732.787)
2 Utah Lake (646.047)
3 United States Forest Service (584.821)
4 Price, Utah (575.731)
5 Red Deer (469.844)
6 Colorado (453.202)
7 Protected areas of the United States (452.932)
8 Utah (451.928)
9 Western United States (431.971)
10 Utah County, Utah (427.949)
16. Generating Clusters with
Explicit Semantic Analysis (ESA)
Query: ‘forest preserves in Utah’
1 U.S. National Monument (732.787) (26.6, utah) (25.7, forest) (13.6, preserv)
(60.4, utah) (13.6, preserv)
2 Utah Lake (646.047)
3 United States Forest Service (584.821)
4 Price, Utah (575.731)
5 Red Deer (469.844)
6 Colorado (453.202)
7 Protected areas of the United States (452.932)
8 Utah (451.928)
9 Western United States (431.971)
10 Utah County, Utah (427.949)
19. Eliminating Redundant Topics
(102, Colorado)(48, Wilderness)(32, Forest)(28, Mountain)(27, Juan)...
(97, Utah)(37, Canyon)(36, Sandstone)(25, Mountain)(21, Colorado)...
(60, Utah)(40, Canyon)(37, Colorado)(36, Sandstone)(23, Mountain)...
(54, Colorado)(44, Canyon)(36, Sandstone)(22, Utah)(20, Mountain)
(34, Utah)(32, Montana)(30, Colorado)(22, Mountain)(18, Forest)...
(61, Canyon)(32, Colorado)(32, Utah)(26, Trail)(19, Forest)...
...
er o i o n
s
h
ta
u
re
co du s
tto
at v
st
ne
nw s
ri
ra
oo ct i
nc
s
d on
h
n ch
w i l b i t a t s o sio n
ra
n ev ada
d er n ess er
p ar
er o s
u
nevada
tah
k
n
si o
ion
ha
o
at v s
er
ne
s
m ot
n es
v ad
u t ah
p ar k
orb
r
il
d es
a
ut c tr a
a h any w ilr a
ik e
i ls
t
t
on
r es
ar l an
nevada
ch fo
ds
es r e
bik es
cr
va ad a
p ar k ea park
ti o
a
n edv
n ne
al
ne
cr o
e r o s i on
20. Timeline Generation
utahs canyons colorados park mountain lake sandstone
wi
ld
l
er
ai
tr
ne
canyons trail mountain bikes wilderness hikers
ss
rid
er
s
deputy rights motorcyclists atvs
montana bitterroot
21. Timeline Generation
utahs canyons colorados park mountain lake sandstone
wi
ld
l
er
ai
tr
ne
canyons trail mountain bikes wilderness hikers
ss
rid
er
s
deputy rights motorcyclists atvs
montana bitterroot
22. Timeline Generation
utahs canyons colorados park mountain lake sandstone
wi
ld
l
er
ai
tr
ne
canyons trail mountain bikes wilderness hikers
ss
rid
er
s
deputy rights motorcyclists atvs
montana bitterroot
23. Timeline Generation
utahs canyons colorados park mountain lake sandstone
wi
ld
l
er
ai
tr
ne
canyons trail mountain bikes wilderness hikers
ss
rid
er
s
deputy rights motorcyclists atvs
montana bitterroot
27. Word Selection in Timeline
trail bikes bikes
trail
motorcycle
atvs bikes motorcycle
wilderness riders
wilderness atvs
trail trail
vehicles riders
atvs riders
recreation mountain
wilderness
wilderness
trail park
wildlife
park
wildlife colorado
mountain mountain
4
78
28. Word Selection in Timeline
trail trail bikes bikes
bikes motorcycle motorcycle
atvs
atvs
wilderness riders
wilderness
trail trail
vehicles riders
atvs riders
recreation mountain
wilderness wilderness
park
trail
wildlife
park
wildlife colorado
mountain mountain
4
78
29. Prototype Contributions
1. Dynamic algorithm to learn conversation models
2. Conversation discourse models
3. Track thematic changes and idea formation
4. Access prior conversation content in near real time
30. Questions
Tony Bergstrom and Karrie Karahalios
University of Illinois at Urbana-Champaign
{abergst2, kkarahal}@cs.uiuc.edu