Automated phrase generation

• US 5,794,193 A
• Filed: 09/15/1995
• Issued: 08/11/1998
• Est. Priority Date: 09/15/1995
• Status: Expired due to Term

• Current Assignee: Lucent Technologies, Inc. (Nokia Corporation)
• Original Assignee: Lucent Technologies, Inc. (Nokia Corporation)
• Inventors: Gorin, Allen Louis
• Primary Examiner(s): Hudspeth, David R.
• Assistant Examiner(s): Chawan, Vijay B.
• Application Number:
  US08/528,577
• Time in Patent Office: 1,061 Days
• Field of Search: 395/2.52, 395/2.59, 395/2, 395/2.4, 395/2.62, 379/67, 379/76
• US Class Current: 704/250
• CPC Class Codes: G10L 15/1815 Semantic context, e.g. disambiguation of the recognition hypotheses based on word meaning
• CPC Class Codes: G10L 15/1822 Parsing for meaning understanding
• CPC Class Codes: G10L 15/197 Probabilistic grammars, e.g. word n-grams

Abstract

A methodology for automated task selection is provided, where the selected task is identified in natural speech of a user making such a selection. A set of meaningful phrases are determined by a grammatical inference algorithm which operates on a predetermined corpus of speech utterances, each such utterance being associated with a specific task objective, and wherein each utterance is marked with its associated task objective. Each meaningful phrase developed by the grammatical inference algorithm can be characterized as having both a Mutual Information value and a Salience value (relative to an associated task objective) above a predetermined threshold.

Claims

1. A method for automated determination of a classification parameter for a selected task, where said selected task is expressed in natural speech of a user, comprising the steps of:

providing a database of speech utterances, each said utterances being characterized as directed to ones of a predetermined set of tasks;

forming from ones of a plurality of said speech utterances a set of speech phrases, each said phrase containing at least one word;

determining a likelihood measure for co-occurrence of constituent words in each of said phrases;

selecting from said set of phrases a subset thereof having a value of said likelihood measure exceeding a predetermined threshold;

determining a significance measure relative to a specified one of said predetermined set of tasks for phrases in said selected subset of said phrases;

selecting from said selected subset of said phrases a set of meaningful phrases having a value of said significance measure exceeding a predetermined threshold, said meaningful phrases constituting said classification parameter.

2. The method for automated determination of a classification parameter of claim 1 wherein said likelihood measure is manifested as a mutual information measure.

3. The method for automated determination of a classification parameter of claim 1 wherein said significance measure is manifested as a salience measure.

4. The method for automated determination of a classification parameter of claim 3 wherein said salience measure is represented as a conditional probability of said task being selected given said speech phrase, said conditional probability being a highest value in a distribution of said conditional probabilities over said set of predetermined tasks.

5. The method for automated determination of a classification parameter of claim 1 wherein said speech utterances are generated in response to a query of a form "How may I help you?".

6. The method for automated determination of a classification parameter of claim 1 wherein each of said speech utterances is labelled with a one of said predetermined set of tasks to which each said utterance is directed.

7. The method for automated determination of a classification parameter of claim 1 wherein said first selecting step and all subsequent steps are iteratively repeated for speech phase of size n words, where n is an integer in a range 1≦

n≦

N, N being selected by an implementer of said method.

8. A method for establishing a classification relationship between at least one speech phrase and a one of a predetermined set of task objectives, wherein each said speech phrase is formed from a one of a known corpus of speech utterances and each said speech utterance in said corpus is related to a one of said predetermined set of task objectives, said method comprising the steps of:

generating a plurality of said speech phrases, each having a predetermined number of words, from said corpus;

evaluating a mutual information measure and a salience measure for each said generated speech phrase relative to a given one of said task objectives;

selecting a portion of said plurality of said speech phrases having said mutual information measure and said salience measure above predetermined thresholds;

generating from said corpus a second plurality of said speech phrases using said selected portion as a base, wherein each said speech phrase in said second plurality contains at least one additional word relative to said number of words comprising said speech phrases generated by the initial generating step;

evaluating a mutual information measure and a salience measure for each said speech phrase in said second plurality of said speech phrases relative to said given one of said task objectives;

selecting a portion of said second plurality of said speech phrases having said mutual information measure and said salience measure above predetermined thresholds;

iteratively repeating the immediately preceding generating, evaluating and selecting steps until a set of speech phrases having a desired relationship with said given one of said task objectives is selected.

9. The method for establishing a classification relationship of claim 8 wherein said salience measure is represented as a conditional probability of said task being selected given said speech phrase, said conditional probability being a highest value in a distribution of said conditional probabilities over said set of predetermined tasks.

10. The method for establishing a classification relationship of claim 8 wherein each of said speech utterances is labeled with a one of said predetermined set of tasks to which each said utterance is directed.

11. The method for establishing a classification relationship of claim 8 wherein said speech utterances are generated in response to a query of the form "How may I help you?".