MAS.632 Conversational Computer Systems - MIT OpenCourseWare

206 VOIC([ OMMUNICATION WITH COMPUTERS
input words: ABCD parser candidates: ABCD
ABC ACD BCD . .
AB AC AD BC , .
A B C D
Figure 9.17 The Conversational Desktop attempted to parse all substrings
of the input utterance. Each substring was evaluated, and those
which could be parsed were candidates for selection.
Because any word reported by a speech recognizer might be a spurious insertion
error, all substrings of the input words were parsed as shown in Figure 9.17.
If any words were recognized correctly, one of the substrings would contain
exactly all these words without the insertions. All of the substrings were submitted
to the parser, which had to reject those strings containing grammatically correct
insertions, but accept the substring that represented the original sentence
minus the insertion.
Rejection errors caused incomplete sentences and resulted in fragments of syntactically
correct tokens. To cope with such errors, the grammar was modified to
describe sentence fragments as well as complete sentences, e.g., the word
sequence "Barry Friday at two o'clock" is a well-formed fragment from a sentence
about scheduling a meeting. This parsing strategy also handled those substitution
errors resulting in semantically incongruous sentences; if one of the substrings
contained exactly the correct words, then the resulting sentence fragment
would be accepted by the parser's rules for fragments.
These parsing strategies had to be combined because any type of error could
occur in each sentence. Because of the fragment rules, multiple substrings from
each input sequence would usually be accepted by the parser; the next stage of
analysis selected the best of these. Evaluation of "best" was based on empirically
derived weightings ofthe following.
* Completeness: Assuming that the user spoke well-formed sentences,
a successfully parsed complete sentence was more likely spoken
than was a sentence fragment.
* Number of words: A greater number of words in an input string
was preferred because recognition results were more likely to be correct
than erroneous. This judgment was also based on the assumption
that the user spoke sentences based only on words in the
recognizer's vocabulary. If both "ABC" and "AB"could be parsed,
"ABC" would receive preference frm this metric.
* Sequences of words: Because of the difficulty of word endpoint
detection, connected recognizers tend to produce runs of correct or
erroneous results. If the endpoint of one word is incorrectly identified,
not only is that word likely to be misrecognized but also the following
word. The scoring metric favored substrings that included