Contextual filtering

• US 7,379,870 B1
• Filed: 02/03/2005
• Issued: 05/27/2008
• Est. Priority Date: 02/03/2005
• Status: Expired due to Fees

• Current Assignee: HRL Laboratories LLC (The Boeing Co.)
• Original Assignee: HRL Laboratories LLC (The Boeing Co.)
• Inventors: Kadambe, Shubha, Belvin, Robert
• Primary Examiner(s): MCFADDEN, SUSAN IRIS
• Application Number:
  US11/051,747
• Time in Patent Office: 1,209 Days
• Field of Search: 704/255
• US Class Current: 704/255
• CPC Class Codes: G10L 15/18 using natural language modelling

Abstract

The present invention relates to techniques for contextual filtering for improving an output of a speech recognizer. The techniques comprise receiving a representation of a speech utterance into a shallow parser component in the form of a first word lattice including a set of potential language matches for the utterance, where the language matches include at least one word; the parser component operative for receiving and analyzing the first word lattice to produce a second word lattice, and including a filter that assigns a probability match for a potential word match in relation to another potential word match in the utterance based on a particular filter-specific criteria; and outputting the probability matches from the shallow parser component as a portion of a second word lattice for further processing to determine most likely combinations of words present in the speech utterance.

Claims

1. A contextual filter for use within a computer-based speech recognition system comprising:

a shallow parser input for receiving a representation of an utterance into a shallow parser component in a form of a first word lattice, the first word lattice including a set of potential textual word sequence matches for the utterance, where the textual word sequence matches include at least one word;

the shallow parser component operative for receiving and analyzing the first word lattice to produce a second word lattice, and including a filter selected from a group consisting of a sub-categorization licensing filter, a predication licensing filter, an adverbial licensing filter, and a discourse licensing filter, where the filter assigns a probability match for a potential word match in relation to another potential word match in the utterance based on a particular filter-specific criteria; and

a shallow parser output for outputting the probability matches from the shallow parser component as a portion of a second word lattice for further processing to determine most likely combinations of words present in the utterance.

2. A contextual filter as set forth in claim 1, wherein the sub-categorization licensing filter is configured to use co-occurrence restrictions found in syntactic lexicons to provide information as to what lexical categories may occur as a grammatical complement of a particular word, and using the lexical categories to determine a likelihood of a match for a given vocabulary item, given the particular word.

3. A contextual filter as set forth in claim 2, wherein in analyzing the first word lattice to produce the second word lattice, the sub-categorization licensing filter is configured to remove strings of text that fall below a predetermined threshold likelihood of match.

4. A contextual filter as set forth in claim 1, wherein the predication licensing filter is configured to identify a relation between a subject and a predicate to determine a likelihood of a match for given vocabulary items.

5. A contextual filter as set forth in claim 1, wherein in analyzing the first word lattice to produce the second word lattice, the predication licensing filter is configured to remove strings of text that fall below a predetermined threshold likelihood of match.

6. A contextual filter as set forth in claim 1, wherein the adverbial licensing filter is configured to identify a relation between modifying or reference-restricting phrases and phrases they modify to determine a likelihood of a match for given vocabulary items.

7. A contextual filter as set forth in claim 1, wherein in analyzing the first word lattice to produce the second word lattice, the adverbial licensing filter is configured to remove strings of text that fall below a predetermined threshold likelihood of match.

8. A contextual filter as set forth in claim 1, wherein the discourse licensing filter is configured to identify a relation between phrases or sentences in a coherent discourse to determine a likelihood of a given match for given vocabulary items.

9. A contextual filter as set forth in claim 1, wherein in analyzing the first word lattice to produce the second word lattice, the discourse licensing filter is configured to remove strings of text that fall below a predetermined threshold likelihood of match.

10. A contextual filter as set forth in claim 1, further comprising:

an acoustic model input for receiving an utterance into an acoustic model component and producing a phoneme lattice for further processing into the word lattice;

the acoustic model component operative for receiving the utterance and analyzing the utterance to produce a phoneme lattice that corresponds to the utterance; and

an acoustic model output for outputting the phoneme lattice for further processing.

11. A contextual filter as set forth in claim 10, further comprising:

a phonotactics/pronunciation model input for receiving the phoneme lattice into a phonotactics/pronunciation model component;

the phonotactics/pronunciation model component operative for receiving the phoneme lattice and converting the phoneme lattice into a first word lattice by applying phonotatic constraints to the phoneme lattice, thereby converting the phoneme lattice into the first word lattice; and

a phonotactics/pronunciation model output for outputting the first word lattice for further processing.

12. A contextual filter as set forth in claim 1, further comprising:

an N-gram language model input for receiving the second word lattice into an N-gram language model component;

the N-gram language model component operative for receiving the second word lattice and converting the second word lattice into an N-best list by applying a language model component that outputs an N-best possible word string for a given utterance that have N highest likelihood values; and

an N-gram language model output for outputting the N-best list from the N-gram language model component for further processing.

13. A contextual filter as set forth in claim 12, further comprising:

a full parser input for receiving the N-best list into a full parser component;

the full parser component operative for receiving the N-best list and converting the N-best list into a parse structure through at least one item selected from a group consisting of a syntactic parser and a semantic parser; and

a full parser output for outputting the parse structure from the full parser component.

14. A contextual filter as set forth in claim 1, wherein in analyzing the first word lattice to produce the second word lattice, and when more than one filter is included in the contextual filter, the contextual filter is configured to multiply the probability match between each filter to create a cumulative probability match, and is further configured to remove strings of text that fall below a predetermined threshold cumulative probability match.

15. A method for contextual filtering, the method operating on a computer system including a processor, an input coupled with the processor for receiving an input, and an output coupled with the processor for producing an output, the method comprising acts of:

receiving a representation of an utterance into a shallow parser component in the form of a first word lattice including a set of potential textual word sequence matches for the utterance, where the textual word sequence matches include at least one word;

receiving and analyzing the first word lattice to produce a second word lattice utilizing a filter selected from a group consisting of a sub-categorization licensing filter, a predication licensing filter, an adverbial licensing filter, and a discourse licensing filter, where the filter assigns a probability match for a potential word match in relation to another potential word match in the utterance based on a particular filter-specific criteria; and

outputting the probability matches from the shallow parser component as a portion of a second word lattice for further processing to determine most likely combinations of words present in the utterance.

16. A method for contextual filtering as set forth in claim 15, wherein the act of receiving and analyzing the first word lattice to produce a second word lattice further comprises an act of configuring the sub-categorization licensing filter to use co-occurrence restrictions found in syntactic lexicons to provide information as to what lexical categories may occur as a grammatical complement of a particular word, and using the lexical categories to determine a likelihood of a match for a given vocabulary item, given the particular word.

17. A method for contextual filtering as set forth in claim 16, wherein the act of receiving and analyzing the first word lattice to produce the second word lattice, further comprises an act of configuring the sub-categorization licensing filter to remove strings of text that fall below a predetermined threshold likelihood of match.

18. A method for contextual filtering as set forth in claim 15, wherein the act of receiving and analyzing the first word lattice to produce a second word lattice further comprises an act of configuring the predication licensing filter to identify a relation between a subject and a predicate to determine a likelihood of a match for given vocabulary items.

19. A method for contextual filtering as set forth in claim 15, wherein the act of receiving and analyzing the first word lattice to produce the second word lattice, further comprises an act of configuring the predication licensing filter to remove strings of text that fall below a predetermined threshold likelihood of match.

20. A method for contextual filtering as set forth in claim 15, wherein the act of receiving and analyzing the first word lattice to produce a second word lattice further comprises an act of configuring the adverbial licensing filter to identify a relation between modifying or reference-restricting phrases and phrases they modify to determine a likelihood of a match for given vocabulary items.

21. A method for contextual filtering as set forth in claim 15, wherein the act of receiving and analyzing the first word lattice to produce the second word lattice, further comprises an act of configuring the adverbial licensing filter to remove strings of text that fall below a predetermined threshold likelihood of match.

22. A method for contextual filtering as set forth in claim 15, wherein the act of receiving and analyzing the first word lattice to produce a second word lattice further comprises an act of configuring the discourse licensing filter to identify a relation between phrases or sentences in a coherent discourse to determine a likelihood of a given match for given vocabulary items.

23. A method for contextual filtering as set forth in claim 15, wherein the act of receiving and analyzing the first word lattice to produce the second word lattice, further comprises an act of configuring the discourse licensing filter to remove strings of text that fall below a predetermined threshold likelihood of match.

24. A method for contextual filtering as set forth in claim 15, further comprising acts of:

receiving an utterance into an acoustic module;

analyzing the utterance to produce a phoneme lattice that corresponds to the utterance; and

outputting the phoneme lattice from the acoustic model component for further processing into the first word lattice.

25. A method for contextual filtering as set forth in claim 24, further comprising acts of:

receiving the phoneme lattice into a phonotactics/pronunciation module;

converting the phoneme lattice into a first word lattice by applying phonotatic constraints to the phoneme lattice, thereby converting the phoneme lattice into the first word lattice; and

outputting the first word lattice for further processing by the parser component.

26. A method for contextual filtering as set forth in claim 15, further comprising acts of:

receiving the second word lattice into an N-gram language module;

converting the second word lattice into an N-best list by applying a language model component that outputs an N-best possible word string for a given utterance that have N highest likelihood values; and

outputting the N-best list from the N-gram language model component for further processing.

27. A method for contextual filtering as set forth in claim 26, further comprising acts of:

receiving the N-best list into a full parser component;

converting the N-best list into a parse structure through at least one item selected from a group consisting of a syntactic parser and a semantic parser; and

outputting the parse structure from the full parser component.

28. A method for contextual filtering as set forth in claim 25, wherein the act of receiving and analyzing the first word lattice to produce the second word lattice, further comprises an act of configuring the contextual filter to multiply the probability match between each filter to create a cumulative probability match, and to remove strings of text that fall below a predetermined threshold cumulative probability match.

29. A computer program product for contextual filtering, comprising computer instruction means stored on a computer readable medium, for:

receiving a representation of an utterance into a shallow parser component in a form of a first word lattice including a set of potential textual word sequence matches for the utterance, where the textual word sequence matches include at least one word;

receiving and analyzing the first word lattice to produce a second word lattice utilizing a filter selected from a group consisting of a sub-categorization licensing filter, a predication licensing filter, an adverbial licensing filter, and a discourse licensing filter, where the filter assigns a probability match for a potential word match in relation to another potential word match in the utterance based on a particular filter-specific criteria; and

outputting the probability matches from the shallow parser component as a portion of a second word lattice for further processing to determine most likely combinations of words present in the utterance.

30. A computer program product for contextual filtering at set forth in claim 29, wherein the computer instruction means for receiving and analyzing the first word lattice to produce a second word lattice further comprises computer instruction means for configuring the sub-categorization licensing filter to use co-occurrence restrictions found in syntactic lexicons to provide information as to what lexical categories may occur as a grammatical complement of a particular word, and using the lexical categories to determine a likelihood of a match for a given vocabulary item, given the particular word.

31. A computer program product for contextual filtering at set forth in claim 30, wherein the computer instruction means for receiving and analyzing the first word lattice to produce the second word lattice further comprises a computer instruction means for causing the sub-categorization licensing filter to remove strings of text that fall below a predetermined threshold likelihood of match.

32. A computer program product for contextual filtering at set forth in claim 29, wherein the computer instruction means for receiving and analyzing the first word lattice to produce a second word lattice further comprises a computer instruction means for configuring the predication licensing filter to identify a relation between a subject and a predicate to determine a likelihood of a match for given vocabulary items.

33. A computer program product for contextual filtering at set forth in claim 29, wherein the computer instruction means for receiving and analyzing the first word lattice to produce the second word lattice, further comprises a computer instruction means for causing the predication licensing filter to remove strings of text that fall below a predetermined threshold likelihood of match.

34. A computer program product for contextual filtering at set forth in claim 29, wherein the computer instruction means for receiving and analyzing the first word lattice to produce a second word lattice further comprises computer instruction means for configuring the adverbial licensing filter to identify a relation between modifying or reference-restricting phrases and phrases they modify to determine a likelihood of a match for given vocabulary items.

35. A computer program product for contextual filtering at set forth in claim 29, wherein the computer instruction means for receiving and analyzing the first word lattice to produce the second word lattice, further comprises a computer instruction means for causing the adverbial licensing filter to remove strings of text that fall below a predetermined threshold likelihood of match.

36. A computer program product for contextual filtering at set forth in claim 29, wherein the computer instruction means for receiving and analyzing the first word lattice to produce a second word lattice further comprises computer instruction means for configuring the discourse licensing filter to identify a relation between phrases or sentences in a coherent discourse to determine a likelihood of a given match for given vocabulary items.

37. A computer program product for contextual filtering at set forth in claim 29, wherein the computer instruction means for receiving and analyzing the first word lattice to produce the second word lattice, further comprises a computer instruction means for causing the discourse licensing filter to remove strings of text that fall below a predetermined threshold likelihood of match.

38. A computer program product for contextual filtering at set forth in claim 29, further comprising computer instruction means for:

receiving an utterance into an acoustic module;

analyzing the utterance to produce a phoneme lattice that corresponds to the utterance; and

outputting the phoneme lattice from the acoustic model component for further processing into the first word lattice.

39. A computer program product for contextual filtering at set forth in claim 38, further comprising computer instruction means for:

receiving the phoneme lattice into a phonotactics/pronunciation module;

converting the phoneme lattice into a first word lattice by applying phonotatic constraints to the phoneme lattice, thereby converting the phoneme lattice into the first word lattice; and

outputting the first word lattice for further processing by the parser component.

40. A computer program product for contextual filtering at set forth in claim 29, further comprising computer instruction means for:

receiving the second word lattice into an N-gram language module;

converting the second word lattice into an N-best list by applying a language model component that outputs an N-best possible word string for a given utterance that have N highest likelihood values; and

outputting the N-best list from the N-gram language model component for further processing.

41. A computer program product for contextual filtering at set forth in claim 40, further comprising computer instruction means for:

receiving the N-best list into a full parser component;

converting the N-best list into a parse structure through at least one item selected from a group consisting of a syntactic parser and a semantic parser; and

outputting the parse structure from the full parser component.

42. A computer program product for contextual filtering at set forth in claim 29, wherein the computer instruction means for receiving and analyzing the first word lattice to produce the second word lattice, further comprises a computer instruction means for causing the contextual filter to multiply the probability match between each filter to create a cumulative probability match, and to remove strings of text that fall below a predetermined threshold cumulative probability match.