Graphics Reference
In-Depth Information
is only aware of actions 1280 and 640 and only explores shorter STWs
for patterns where the lowest available STW is considered the best.
6. Results
To reiterate, there are three conditions: Artificial, Single person, and
10 people. First we will answer the question of whether the system
is learning; then we will look at the above dependent measures in
more detail.
6.1 Is the system learning?
The system showed significant learning effects for the Artificial
condition, both for reaction time (simple regression
F
= 12.83;
p
<
0.0005) and overlaps (simple regression
F
= 10.41;
p
< 0.0047). The
system also showed significant learning effects for the 10-person
condition, for reaction time (see Table 2), and overlaps (see Table
3). Although an 89 msec gain in STW may seem small, it makes a
big qualitative difference for most average dialogue participants,
essentially changing an automatic dialogue system from being
obviously inadequate and sometimes annoyingly slow to not being
so. The system starts each interview with previous learning and thus
optimal STW based on another person's prosody patterns instead of
beginning with a “safe” 1-2 second STW. To shorten this previous
optimal STW, at the same time as overlaps drop from 24% to 10%,
shows that the agent is learning new skills on the fly, becoming
increasingly more “polite” (efficient and cooperative) by improving
its reaction time and speech overlap performance between- as well
as within-interviews.
Table 2.
Paired one-tail
t
-test: Interviewing 10 consecutive people.
Turn
Observation (
N
)
Mean
St.Dev
Turn 1-15
10
655 msecs
137.25
Turn 16-30
10
566 msecs
73.83
T
-value = 2.46,
P
-value = 0.018, DF = 9
Table 3.
Paired one-tail
t
-test: Overlaps when interviewing 10 consecutive people.
Turn
Observation (
N
)
Mean
St.Dev
Turn 1-15
10
0.24
0.11
Turn 16-30
10
0.10
0.09
T
-value = 4.16,
P
-value = 0.0012, DF = 10
Search WWH ::
Custom Search