( current & future work ) explicit confirmation implicit confirmation unplanned implicit...
-
Upload
ashley-lowe -
Category
Documents
-
view
212 -
download
0
Transcript of ( current & future work ) explicit confirmation implicit confirmation unplanned implicit...
( current & future work ) explicit
confirmation
implicitconfirmation
unplanned implicitconfirmation
request
constructing accurate beliefs in spoken dialog systems
Dan Bohus, Alexander I. RudnickyComputer Science Department, Carnegie Mellon University
1 abstract
phone-based, mixed initiative system for conference room reservations access to live schedules for 13 rooms in 2 buildings size, location, a/v equipment
Roomline
( current & future work )We propose a data-driven approach for constructing more accurate
beliefs over concept values in spoken dialog systems by integrating information across multiple turns in the conversation. The approach bridges existing work in confidence annotation and correction detection and provides a unified framework for belief updating. It significantly outperforms heuristic rules currently used in most spoken dialog systems.
3 user response analysis a k-hypotheses + other
4 models. results
5 conclusion
2 problem
3 dataset
b estimated impact on task success
c using information from n-best lists
user study with the RoomLine spoken dialog system 46 participants (1st time users) 10 scenario-based interactions each 449 dialogs 8278 turns
corpus transcribed and annotated
responses to explicit confirmations
Yes No Other
Correct (1159)
94% [93%]
0% [0%] 5% [7%]
Incorrect (279)
1% [6%] 72% [57%]
27% [37%]
responses to implicit confirmations
Yes No Other
Correct (554)
30% [0%] 7% [0%] 63% [100%]
Incorrect (229)
6% [0%] 33% [15%]
61% [85%]
how do users respond to correct and incorrect confirmations?
explicit confirmation
User correc
ts
User does not
correct
Correct 0 1159
Incorrect
250 29
implicit confirmation
User correc
ts
User does not
correct
Correct 2 552
Incorrect
111 118
users interact strategically
~ correct later
correct
later
~critical
55 2
critical 14 47
how often users correct the system?
31.15
8.41
3.57 2.71
30%
20%
10%
0%
30.40
23.37
16.15 15.33
30%
20%
10%
0%
15.4014.36
12.6410.37
20%
10%
0%
explicit confirmation implicit confirmation unplanned implicit confirmation
proposed a data-driven approach for constructing more accurate beliefs in task-oriented spoken dialog systems
bridge insights from detection of misunderstandings and corrections into a unified belief updating framework
model significantly outperforms heuristics currently used in most spoken dialog systems
initial initial confidence score of top hypothesis, # of initial hypotheses, concept type (bool / non-bool), concept identity;
system action indicators describing other system actions in conjunction with current confirmation;
user response
acoustic / prosodic
acoustic and language scores, duration, pitch (min, max, mean, range, std.dev, min and max slope, plus normalized versions), voiced-to-unvoiced ratio, speech rate, initial pause;
lexical number of words, lexical terms highly correlated with corrections (MI);
grammatical
number of slots (new, repeated), parse fragmentation, parse gaps;
dialog dialog state, turn number, expectation match, new value for concept, timeout, barge-in.
evaluation initial(error rate in system beliefs before the update)
heuristic(error rate in system beliefs after the update – using the heuristic update rules)
proposed model(error rate of the proposedlogistic model tree)
oracle(oracle error rate)
features
model
Confupd(thC) ← M (Confinit(thC), SA(C), R) logistic model tree [one for each system action]
1-level deep, root splits on answer-type (YES / NO / other) leaves contain stepwise logistic regression models
sample efficient, feature selection good probability outputs (minimize cross entropy between model predictions and reality)
10%
0%
20%
30%
16.17%
5.52%
30.83%
6.06%7.86%
10%
20%
0%
26.16%
17.56%
30.46%
21.45%22.69%
0%
12%
4%
8%
15.15%
10.72%
15.49%
14.02%12.95%
4%
0%
8%
12%9.49%
6.08%
98.14%
9.64%
45.03%
19.23%
80.00%
25.66%
20%
0%
40%
unexpectedupdates
initial
heuristic
proposed model(basic feature set)
proposed model(basic + priors)
oracle
As a prerequisite for increased robustness and making better decisions, dialog systems must be able to accurately assess the reliability of the information they use. Typically, recognition confidence scores provide an initial assessment for the reliability of the information obtained from the user. Ideally, a system should leverage information available in subsequent turns to update and improve
the accuracy of it’s beliefs.
belief updating problem:given an initial belief over a concept Belieft(C), a system
action SA(C) and a user response R, compute the updated belief Belieft+1(C)
S: starting at what time do you need the room?U: [STARTING AT TEN A M / 0.45]
starting at ten a.m.start-time = {10:00 / 0.45}
S: did you say you wanted the room starting at ten a.m.?U: [GUEST UNTIL ONE / 0.89]
yes until noon
start-time = {?}
explicit confirmation (correct value)
S: for when do you need the room?U: [NEXT THURSDAY / 0.75]
next Thursdaydate = {2004-08-26 / 0.75}
S: a room for Thursday, August 26th … starting at what
time do you need the room?U: [FIVE TO SEVEN P_M / 0.58]
five to seven p.m.
date = {?}
implicit confirmation (correct value)
S: how may I help you?U: [THREE TO RESERVE A ROOM / 0.65]
I’d like to reserve a roomstart-time = {15:00 / 0.65}
S: starting at three p.m. … for which day do you need the
conference room?U: [CAN YOU DETAILS TIME / NONUNDER.(0.0)]
I need a different time
start-time = {?}
implicit confirmation (incorrect value)
belief representation k hypotheses + other multinomial generalized linear model
system actions: all actions explicit confirmation implicit confirmation unplanned impl. confirmation request [system asks for the value for a concept] unexpected update [system received a value for a concept, without asking for it, e.g. as a result of a misrecognition or the user over-answering or attempting a topic shift]
features added prior information on concepts priors constructed manually
belief representation: most accurately: probability distribution over the set of possible values but: system is not likely to “hear” more than 3 or 4 conflicting values
in our data, the maximum number of hypotheses for a concept accumulated through interaction was 3; the system heard more than 1 hypothesis for a concept in only 6.9% of cases
compressed belief representation: k hypotheses + other for now, k = 1: top hypothesis + other [see current and future work for extensions]
for now, only updates after system confirmation actions
given an initial confidence score for the top hypothesis h for a concept C - Confinit(thC), construct an updated confidence score for the hypothesis
h - Confupd(thC) - in light of the system confirmation action SA(C) and
the follow-up user response R
compressed belief updating problem:
how does the accuracy of the belief updating model affect task success? relates the accuracy of the belief updates to overall task success through a logistic regression model accuracy of belief updates: measured as AVG-LIK of the correct hypothesis word-error-rate acts as a co-founding factor
model: P(Task Success=1) ← α + β•WER + γ•AVG-LIK fitted model using 443 data-points (dialog sessions) β, γ capture the impact of WER and AVG-LIK on overall task success
0 20 40 60 80 1000
0.2
0.4
0.6
0.8
1
0 20 40 60 80 1000
0.2
0.4
0.6
0.8
1
0 20 40 60 80 100
0.2
0.4
0.6
0.8
0.0
1.0
0.2
0.4
0.6
0.8
0.0
1.0
0 20 40 60 80 100
word-error-rate word-error-rate
pro
bab
ility
of
task
su
ccess
avg.lik. = 0.5
avg.lik. = 0.6
avg.lik. = 0.7
avg.lik. = 0.8
avg.lik. = 0.9
current heuristic
proposed model
avg.lik. = 0.5
avg.lik. = 0.6
current heuristic
avg.lik. = 0.7
proposed model
avg.lik. = 0.8
avg.lik. = 0.9
avera
ge w
ord
-err
or
rate
avera
ge w
ord
-err
or
rate
natives non-natives
pro
bab
ility
of
task
su
ccess
currently: using only the top hypothesis from the recognizer next: extract more information from n-best list or lattices