System and method for generating natural language phrases from user utterances in dialog systems

• US 8,874,443 B2
• Filed: 08/27/2008
• Issued: 10/28/2014
• Est. Priority Date: 08/27/2008
• Status: Active Grant

• Current Assignee: Robert Bosch GmbH
• Original Assignee: Robert Bosch GmbH
• Inventors: Weng, Fuliang, Stoia, Laura, Hu, Junling, Feng, Zhe, Cao, Junkuo
• Primary Examiner(s): Dorvil, Richemond
• Assistant Examiner(s): ADESANYA, OLUJIMI A
• Application Number:
  US12/199,520
• Publication Number:
  US 20100057463A1
• Time in Patent Office: 2,253 Days
• Field of Search: 704/257
• US Class Current: 704/257
• CPC Class Codes: G06F 40/20 Natural language analysis semantic analysis of natural language G06F40/30
• CPC Class Codes: G06F 40/40 Processing or translation of natural language natural language analysis G06F40/20; semantic analysis G06F40/30
• CPC Class Codes: G10L 15/22 Procedures used during a speech recognition process, e.g. man-machine dialogue

Abstract

Embodiments of a dialog system that employs a corpus-based approach to generate responses based on a given number of semantic constraint-value pairs are described. The system makes full use of the data from the user input to produce dialog system responses in combination with a template generator. The system primarily utilizes constraint values in order to realize efficiencies based on the more frequent tasks performed in real dialog systems although rhetorical or discourse aspects of the dialog could also be included in a similar way, that is, labeling the data with such information and performing a training process. The benefits of this system include higher quality user-aligned responses, broader coverage, faster response time, and shorter development cycles.

Claims

1. A method of generating responses in a dialog system in response to user input in a current conversation session, comprising:

collecting original user phrase data comprising the user input from one or more past conversation sessions between a first user only and the dialog system to form phrase models in a training phase using the first user'"'"'s own phrases for use by a response generator in response to a query by the first user;

using the phrase models at a semantic level to use wordings in the response that utilize the first user'"'"'s own phrases and to identify constraint-carrying phrases from the first user'"'"'s utterances;

determining constraints expressed in each user utterance in the past conversation sessions using one or more constraint identifiers;

forming a pool of user examples instructing the system on how to lexicalize the constraints to form lexicalized constraints;

extracting one or more extracted constraints, each comprising a constraint-value pair, from the user input;

generalizing the extracted constraints into phrase patterns using the lexicalized constraints to produce general phrase patterns;

constructing a response database consisting of the original phrase data and the general phrase patterns;

transmitting a set of filled constraints to the response generator component from the user input; and

selecting at least one phrase from the response database that optimally matches the set of filled constraints to form a response utilizing the first user'"'"'s own phrases.

2. The method of claim 1 wherein the one or more conversation sessions comprises a series of conversation turns, and wherein selecting the at least one phrase from the response database comprises using a most recent utterance from the first user.

3. The method of claim 2 further comprising:

utilizing a chunking-based approach to extract phrases used by the first user in a constraint specification; and

updating the extracted phrases with values retrieved from the response database.

4. The method of claim 1 wherein the one or more conversation sessions comprises at least one previous session between the first user and the dialog system, and not including the current session user input.

5. The method of claim 4 further comprising:

utilizing one of a parse-tree based approach or a chunking-based approach to extract phrases used by the first user in a constraint specification and form a pool of examples of how to lexicalize constraints in the user input; and

generalizing the constraints into phrase patterns by replacing constraints with corresponding semantic class names.

6. The method of claim 5 wherein the parse-tree based approach structurally annotates all user utterances as parse-trees.

7. The method of claim 5 wherein the chunking-based approach tags all first user utterances using a part-of-speech tagger, and parsed by one of a shallow parser or deep structure parser.

8. The method of claim 1 further comprising:

determining whether or not the constraint and value pairs of the filled constraints are present in the response database; and

configuring the response generator to use the corresponding phrase in the response database if the constraint and value pairs of the filled constraints are present in the response database.

9. The method of claim 1 further comprising:

determining whether or not the constraint and value pairs of the filled constraints are present in the response database; and

configuring the response generator to use a corresponding phrase in the response database that overlaps the user input constraint and value pairs of the filled constraints to as high a degree if possible, if the values of the constraints are different between the filled constraints of the database and the values of the constraints in the user input.

10. The method of claim 1 further comprising:

determining whether or not the constraint and value pairs of the filled constraints are present in the response database; and

creating constraint patterns from existing or generated patterns if the constraint and value pairs of the filled constraints are not present in the response database.

11. The method of claim 1 further comprising:

indexing, in the training phase, each sentence in the past conversation sessions by a speech act, action, and constraint-set triple; and

performing a generalization process to obtain sentence templates as a back-off model to enable processing of units in a sentence that come from either a phrase or a sentence uttered by the first user, wherein the generalization process identifies constraint-sets present in the past conversation sessions but having different values, and wherein the back-off model utilizes the lexicalized constraints to align system variability due to the different values, with the original phrase data.

12. A method of generating phrases in a dialog system in response to user input during one or more turns of a present conversation session between a first user and the dialog system, comprising:

collecting, during a training phase, a corpus of user data from one or more past conversation sessions between the first user only and the dialog system to form phrase models using the first user'"'"'s own phrases for use by a response generator component in response to a query by the first user;

determining constraints expressed in each user utterance in the past conversation sessions using one or more constraint identifiers;

forming a pool of user examples instructing the system on how to lexicalize the constraints to form lexicalized constraints;

using the phrase models at a semantic level to use wordings in the response that utilize the first user'"'"'s own phrases;

constructing phrase models based on the corpus of first user data, wherein the phrase models comprise constraint sets consisting of constraint-value pairs based on the lexicalized constraints;

selecting an optimal constraint set from the corpus of user data in response to a constraint within the user input; and

generating a sentence for output from the response generator component incorporating the optimal constraint set to form a response utilizing the first user'"'"'s own phrases.

13. The method of claim 12 wherein the step of constructing phrase models further comprises:

identifying phrase expressions from the first user utterances that have constraints;

generalizing identified constraints through a decomposition process; and

creating one or more default phrase expressions for each constraint set.

14. The method of claim 12 wherein the step of selecting an optimal constraint set further comprises:

determining whether or not the constraint and value pairs of the filled constraints are present in the response database;

configuring the response generator component to use the corresponding phrase in the response database if the constraint and value pairs of the filled constraints are present in the response database;

configuring the response generator component to use a corresponding phrase in the response database that overlaps the user input constraint and value pairs of the filled constraints to as high a degree if possible, if the values of the constraints are different between the filled constraints of the database and the values of the constraints in the user input; and

creating constraint patterns from existing or generated patterns if the constraint and value pairs of the filled constraints are not present in the response database.

15. The method of claim 12 wherein the user input comprises one or more sentences, each sentence consisting of a triple comprising a speech act, an action, and a constraint set, and wherein the step of generating a sentence for output from a response generator component further comprises:

indexing each sentence of the user input into a corresponding triple;

splitting the constraint set of the triple into a plurality of smaller subsets if the triple is not present in the corpus;

generating a sequence of triples from the split step;

generate a corresponding sentence for each triple of the sequence of triples based on a defined training model; and

reducing redundancies in the generated corresponding sentences.

16. The method of claim 15 wherein the step of reducing redundancies further comprises:

determining if a noun phrase used in a current sentence is also used in a previous sentence and refers to an identical entity; and

replacing the current noun phrase with a pronoun with the same number, person, and gender if the noun phrase is used in the previous sentence.

17. The method of claim 12 further comprising performing a generalization process to obtain sentence templates as a back-off model to enable processing of units in a sentence that come from either a phrase or a sentence uttered by the first user;

wherein the generalization process identifies constraint-sets present in the past conversation sessions but having different values, and wherein the back-off model utilizes the lexicalized constraints to align system variability due to the different values, with the original phrase data.

18. The method of claim 17 further comprising:

utilizing a chunking-based approach to extract phrases used by the first user in a constraint specification; and

updating the extracted phrases with values retrieved from the response database.

19. The method of claim 12 wherein the corpus of user data from one or more conversation sessions comprising at least one previous session not including the current session user input.

20. The method of claim 19 further comprising:

utilizing one of a parse-tree based approach or a chunking-based approach to extract phrases used by the first user in a constraint specification and form a pool of examples of how to lexicalize constraints in the user input; and

generalizing the constraints into phrase patterns by replacing constraints with corresponding semantic class names.

21. The method of claim 20 wherein the parse-tree based approach structurally annotates all user utterances as parse-trees.

22. The method of claim 20 wherein the chunking-based approach tags all user utterances using a part-of-speech tagger, and parsed by one of a shallow parser or deep structure parser.