Method and apparatus for creating speech recognition grammars constrained by counter examples

• US 5,937,385 A
• Filed: 10/20/1997
• Issued: 08/10/1999
• Est. Priority Date: 10/20/1997
• Status: Expired due to Fees

• Current Assignee: International Business Machines Corporation
• Original Assignee: International Business Machines Corporation
• Inventors: Kambhatla, Nandakishore, Zadrozny, Wlodek Wlodzimierz
• Primary Examiner(s): Hudspeth, David R.
• Assistant Examiner(s): SAX, ROBERT L
• Application Number:
  US08/954,761
• Time in Patent Office: 659 Days
• Field of Search: 704/257, 704/255, 704/251
• US Class Current: 704/257
• CPC Class Codes: G10L 15/183 using context dependencies, e.g. language models
• CPC Class Codes: G10L 15/193 Formal grammars, e.g. finite state automata, context free grammars or word networks

Abstract

An automated system generates and revises grammars for speech recognizers in a speech recognition system. Given an initial grammar, expressed in terms of non-terminals in Backus-Naur Form (BNF) notation, a sentence generator generates a list of all sentences accepted by the grammar. From this list, a corpus of inappropriate or irrelevant sentences which are accepted by the grammar (counter-examples) is identified. A grammar revisor program uses the original grammar and the list of counter examples, to generate a pruned list from which a revised grammar is generated. The revision process is iterated several times either concatenating or merging pairs of non-terminals until the revised grammar is deemed satisfactory in that it accepts as legal only relevant sentences. The revised grammar is used by the speech recognizer, thus reducing errors in the overall system.

Claims

1. A system for creating grammars for a speech recognition application, comprising:

speech grammar repository means for receiving an initial grammar for a speech recognition system and outputting a set of initial grammar rules for said initial grammar;

sentence generator means for generating a list of possible sentences according to said initial grammar rules;

counter-example generator means for analyzing said list of possible sentences and generating a list of counter example sentences comprising sentences which are legal within said initial grammar rules but which are inapplicable for said speech recognition application; and

a grammar revisor means, receiving said initial grammar rules and said list of counter example sentences, for generating from said initial grammar rules and said list of counter example sentences a revised grammar, wherein said counter example sentences are illegal in said revised grammar.

2. A system for creating grammars for a speech recognition application as recited in claim 1 wherein said grammar revisor means comprises:

counter-example remover means for removing said counter example sentences from said list of possible sentences and outputting a pruned list; and

a grammar generator means, receiving said pruned list and outputting said revised grammar.

3. A system for creating grammars for a speech recognition application, comprising:

speech grammar repository means for receiving an initial grammar for a speech recognition system and outputting a set of initial grammar rules for said initial grammar;

sentence generator means for generating a list of possible sentences according to said initial grammar rules;

counter-example generator for analyzing said list of possible sentences and generating a list of counter example sentences comprising sentences which are legal within said initial grammar rules but which are inapplicable for said speech recognition application; and

a grammar revisor means, receiving said initial grammar rules and said list of counter example sentences, for generating therefrom a revised grammar wherein said counter example sentences are illegal for said speech recognition application,

wherein said grammar revisor means comprises;

counter-example remover means for removing said counter-example sentences from said list of possible sentences and outputting a pruned list; and

a grammar generator means, receiving said pruned list and outputting said revised grammar, wherein said counter example sentences are illegal for said speech recognition application,

wherein said grammar generator means comprises;

means for generating a unique non-terminal for each word in said pruned list;

means for computing a base optimization criterion value based on properties of words in said pruned list for said initial grammar;

means for concatenating a pair of non-terminals and computing a first resultant optimization criterion value;

means for merging said pair of non-terminals and computing the second resultant value of said optimization criterion; and

means for adding a new grammar rule to said initial grammar in accordance with an optimal of said first and second optimization criterion values.

4. A system for creating grammars for a speech recognition application as recited in claim 3 wherein said optimization criterion is description length of the grammar.

5. A system for creating grammars for a speech recognition application as recited in claim 4 wherein said initial grammar is expressed in Backus-Naur Form (BNF) notation.

6. A system for creating grammars for a speech recognition application, comprising:

speech grammar repository means for receiving an initial grammar for a speech recognition system and outputting a set of initial grammar rules for said initial grammar;

sentence generator means for generating a list of possible sentences according to said initial grammar rules;

counter-example generator for analyzing said list of possible sentences and generating a list of counter example sentences comprising sentences which are legal within said initial grammar rules but which are inapplicable for said speech recognition application; and

a grammar revisor means, receiving said initial grammar rules and said list of counter example sentences, for generating therefrom a revised grammar wherein said counter example sentences are illegal for said speech recognition application,

wherein said grammar revisor means comprises;

counter-example remover means for removing said counter-example sentences from said list of possible sentences and outputting a pruned list; and

a grammar generator means, receiving said pruned list and outputting said revised grammar, wherein said counter example sentences are illegal for said speech recognition application,

wherein said grammar generator means comprises;

means for generating a unique non-terminal for each word in said pruned list;

means for computing a base value of an optimization criterion based on properties of the words in said pruned list for said initial grammar;

means for concatenating a pair of said non-terminals and computing a resultant value of said optimization criterion and for adding a new rule for said concatenated pair to said initial grammar if said value of criterion is more optimal than said base value of said optimization criterion; and

means for merging said pair of said non-terminals and computing a resultant value of said criterion and for adding a new rule for said merged pair to said initial grammar if said resultant value of optimization criterion is more optimal than said base value of criterion.

7. A system for creating grammars for a speech recognition application as recited in claim 6 wherein said criterion is description length of the grammar.

8. A system for creating grammars for a speech recognition application as recited in 7 wherein said initial grammar is expressed in Backus-Naur Form (BNF) notation.

9. A method for modifying grammars for a speech recognition application, comprising steps of:

creating an initial grammar for a speech recognition system and outputting a set of initial grammar rules for said initial grammar;

generating a list of possible sentences according to said initial grammar rules;

generating a list of counter-example sentences comprising sentences which are legal within said initial grammar rules but which are inapplicable for said speech recognition application; and

generating from said initial grammar rules and said counter-example sentences a revised grammar wherein said counter example sentences are illegal.

10. A method for modifying grammars for a speech recognition application as recited in claim 9 further comprising steps of:

removing said counter-example sentences from said list of possible sentences and outputting a pruned list; and

revising said initial grammar with said pruned list and outputting a revised grammar wherein said counter example sentences are illegal.

11. A method for modifying grammars for a speech recognition application comprising the steps:

creating an initial grammar for a speech recognition system and outputting a set of initial grammar rules for said initial grammar;

generating a list of possible sentences according to said initial grammar rules;

generating a list of counter-example sentences comprising sentences which are legal within said initial grammar rules but which are inapplicable for said speech recognition application;

generating from said initial grammar rules a revised grammar wherein said counter example sentences are illegal;

removing said counter-example sentences from said list of possible sentences and outputting a pruned list;

revising said initial grammar with said pruned list and outputting a revised grammar wherein said counter example sentences are illegal;

generating a unique non-terminal for each word in said pruned list;

computing the base value of an optimization criterion based on properties of the words in said pruned list for said initial grammar;

concatenating a pair of said non-terminals and computing a first resultant value of said optimization criterion;

merging said pair of non-terminals and computing a second resultant value of said optimization criterion; and

adding a new grammar rule to said initial grammar in accordance with an optimal of said first and second resultant value.

12. A method for modifying grammars for a speech recognition application as recited in claim 11, wherein said optimization criterion is description length of the grammar.

13. A method for modifying grammars for a speech recognition application as recited in claim 12 further comprising the step of expressing said initial grammar in Backus-Naur Form (BNF) notation.

14. A method for modifying grammars for a speech recognition application comprising the steps of:

creating an initial grammar for a speech recognition system and outputting a set of initial grammar rules for said initial grammar;

generating a list of possible sentences according to said initial grammar rules;

generating a list of counter-example sentences comprising sentences which are legal within said initial grammar rules but which are inapplicable for said speech recognition application;

generating from said initial grammar rules a revised grammar wherein said counter example sentences are illegal;

removing said counter-example sentences from said list of possible sentences and outputting a pruned list;

revising said initial grammar with said pruned list and outputting a revised grammar wherein said counter example sentences are illegal;

generating a unique non-terminal for each word in said pruned list;

computing a base value of an optimization criterion based on properties of the words in said pruned list for said initial grammar;

concatenating a pair of said non-terminals and computing a resultant value of said criterion and adding a new rule for said concatenated pair to said initial grammar if said resultant value of optimization criterion is more optimal than said base value of optimization criterion; and

merging said pair of said non-terminals and computing a resultant value of said optimization criterion and adding a new rule for said merged pair to said initial grammar if said resultant value of said optimization criterion is more optimal than said base value of optimization criterion.

15. A method for modifying grammars for a speech recognition application as recited in claim 14, wherein said criterion is description length of the grammar.

16. A method for modifying grammars for a speech recognition application as recited in claim 14 further comprising the step of expressing said initial grammar in Backus-Naur Form (BNF) notation.

17. A method for modifying grammars for a speech recognition application comprising the steps of:

creating an initial grammar for a speech recognition system and outputting a set of initial grammar rules for said initial grammar;

generating a list of possible sentences according to said initial grammar rules;

generating a list of counter-example sentences comprising sentences which are legal within said initial grammar rules but which are inapplicable for said speech recognition application;

generating from said initial grammar rules a revised grammar wherein said counter example sentences are illegal;

generating a parse tree comprising a plurality of non-terminals for each of said sentences in said list of counter-example sentences;

adding a new non-terminal for each non-terminal in said parse tree which generates each phrase that said non-terminal generates but excludes phrases in said list of counter-example sentences; and

modifying a START non-terminal in said parse tree wherein said START non-terminal excludes sentences in said list of counter-example sentences.

18. A method for modifying grammars for a speech recognition application comprising the steps of:

creating an initial grammar for a speech recognition system and outputting a set of initial grammar rules for said initial grammar;

generating a list of possible sentences according to said initial grammar rules;

generating a list of counter-example sentences comprising sentences which are legal within said initial grammar rules but which are inapplicable for said speech recognition application;

generating from said initial grammar rules a revised grammar wherein said counter example sentences are illegal;

generating a parse tree comprising a plurality of non-terminals <

N>

for each of said sentences in said list of counter-example sentences;

for each non-terminal <

N >

in said parse tree except a <

<

Start>

>

symbol;

creating a new rule <

N'"'"'>

such that <

N'"'"'>

generates all phrases generated by <

N>

except any phrases which are in said list of counter-example sentences;

creating a new rule <

No>

such that <

No>

generates only phrases which are in said list of counter-example sentences and generated by <

N>

; and

modifying a rule containing said <

<

START>

>

symbol in said parse tree such that said rule containing said <

<

START>

>

symbol ceases to generate said sentences in said list of counter-example sentences.