Digital Signal Processing Reference
In-Depth Information
We found that introducing the next speaker first as a passenger had an
additional effect of eliciting additional questions from the current speaker; for
example:
The next speaker nudged the current speaker when the current speaker
forgot to ask a pertinent question;
When the operator paused the dialog to search for requested information,
the current speaker and the next speaker conferred regarding what to ask
next, as shown in Figure 3-7.
a)
b)
These improvements helped the current speaker to generate appropriate
questions more easily. The operator commented that the speakers in Phase II
were more proactive than Phase I. Analysis of the corpus revealed that the
initial 49 speakers in Phase I required 68 instruction explanations from the
operator, whereas the following 201 speakers in Phase II required only one
instruction explanation.
Figure 3-7.
Example discussion between the current speaker and the next speaker (Phase II).
4.
TRANSCRIPTION AND TAGGING OF SPOKEN
DIALOG CORPUS
Each dialogue was transcribed into text by hand from audio tape. Each
dialog was segmented into individual utterances; the prefixes 'L:' and 'R:'
were used to label operator and speakers utterances, respectively.
Proper nouns in the text were annotated using the bracketted format {
X
name}, where
X
is one of the 5 letters
{A, P, R, S, W}
representing tourist
attractions, places, railway facilities, shops/restaurants, and traffic facilities,
respectively. Traffic facilities include the names of roads, entrances/exits of
