Browse Prior Art Database

Automatic Speech Recognition Enhancement Based on Syntactic and Semantic Cues

IP.com Disclosure Number: IPCOM000107311D
Original Publication Date: 1992-Feb-01
Included in the Prior Art Database: 2005-Mar-21
Document File: 2 page(s) / 92K

Publishing Venue

IBM

Related People

Braden-Harder, L: AUTHOR [+4]

Abstract

Current recognition systems frequently make mistakes during recognition that result in syntactically and semantically incorrect sentences. The technique described here can be used to identify and correct such errors. Syntactic errors can be identified by a syntactic parser. For example, in the sentence (from a speech recognizer).

This text was extracted from an ASCII text file.
This is the abbreviated version, containing approximately 52% of the total text.

Automatic Speech Recognition Enhancement Based on Syntactic and Semantic Cues

       Current recognition systems frequently make mistakes
during recognition that result in syntactically and semantically
incorrect sentences.  The technique described here can be used to
identify and correct such errors.  Syntactic errors can be identified
by a syntactic parser.  For example, in the sentence (from a speech
recognizer).

      The driver tried vainly two six (to fix) my car.
a syntactic parser can determine that this is not a valid syntactic
structure in English.  At this point, commonly confused words
(two/to, six/fix) could be substituted until a parse can be found.
Subsequent processing can identify words that are the correct part of
speech, but incorrect semantically, by suggesting substitutions for
words using online dictionaries or corpora.  For example, in the
sentence (from a speech recognizer).
      The problem was that several police officers had poured
(pulled) over several cars.
dictionary information could be used to determine that cars are more
likely to be pulled over than to be poured over. The entries for pull
over and car in the on-line version of Longman's Dictionary of
Contemporary English include the domain label of TRANSPORT and the
definition of pull over identified vehicle as a typical thing that is
pulled over. Such post-recognizer processing can reliably identify
those sentences with errors and to select more appropriate (but
similar sounding) lexical items.

      Processing would include the following steps given a sentence
generated by a speech recognizer from incoming acoustic signals.
1.   The "recognized" sentence is processed by a syntactic parser.
If a valid syntactic sentence structure can be found, the sentence is
passed to Step 3.
2.   Otherwise, substitutions are tried for words in the sentence
based on a commonly confused word or group of words that is a
different part of speech.  Words near "break points" in the syntactic
structure are tried first (i.e., words that caused the parse to
fail).  If a substitution results in an acceptable syntactic
structure, the sentence is passed to the next step. (Note: for list
items, headings and other items that are typically incomplete
sentences, this step might be skipped.)
3.   The parser (1) was developed to su...