Reactive learning for efficient dialog tree expansion

• US 9,812,127 B1
• Filed: 04/29/2016
• Issued: 11/07/2017
• Est. Priority Date: 04/29/2016
• Status: Active Grant

• Current Assignee: Conduent Business Services, LLC (Conduent, Inc.)
• Original Assignee: Conduent Business Services, LLC (Conduent, Inc.)
• Inventors: Perez, Julien, Monet, Nicolas
• Primary Examiner(s): Godbold, Douglas
• Application Number:
  US15/142,187
• Publication Number:
  US 20170316777A1
• Time in Patent Office: 557 Days
• CPC Class Codes: G06F 40/169 Annotation, e.g. comment data or footnotes
• CPC Class Codes: G06F 40/35 Discourse or dialogue representation
• CPC Class Codes: G10L 15/22 Procedures used during a speech recognition process, e.g. man-machine dialogue

Abstract

A method for generating dialogs for learning a dialog policy includes, for each of at least one scenario, in which annotators in a pool of annotators serve as virtual agents and users, generating a respective dialog tree in which each path through the tree corresponds to a dialog and nodes of the tree correspond to dialog acts provided by the annotators. The generation includes computing a measure of uncertainty for nodes in the dialog tree, identifying a next node to be annotated, based on the measure of uncertainty, selecting an annotator from the pool to provide an annotation for the next node, receiving an annotation from the selected annotator for the next node, and generating a new node of the dialog tree based on the received annotation. A corpus of dialogs is generated from the dialog tree.

Claims

1. A method for generating dialogs and learning a dialog policy for a dialog system, comprising:

for each of at least one scenario, in which annotators in a pool of annotators serve as virtual agents and users, generating a respective dialog tree in which each path through the tree corresponds to a dialog and nodes of the tree correspond to turn of a dialog, the generation comprising, with a processor;

a) computing a measure of uncertainty for nodes in the dialog tree, comprising;

for each of a plurality of nodes, computing a conflict coefficient C_i which quantifies the diversity of its child-node set, as a function of;

2. The method of claim 1, wherein the at least one scenario comprises a plurality of scenarios and the computing of the measure of uncertainty is performed for nodes of the respective dialog trees and the identifying a next node to be annotated is performed using the computed measure of uncertainty for the nodes of all of the plurality of dialog trees.

3. The method of claim 1, wherein the computing of the measure of uncertainty comprises at least one of:

considering the variation in the subtree of each node, and

modeling a current dialog policy for each dialog tree and identifying a node whose subtree is most at variance with the current dialog policy.

4. The method of claim 1, wherein the identifying a next node comprises identifying one of the plurality of nodes whose conflict coefficient is lowest or which is at least lower than other nodes in the plurality of nodes.

5. The method of claim 1, wherein the selecting of the annotator from the pool comprises selecting the annotator from annotators in the pool that have not provided an annotation for the selected node.

6. The method of claim 1, wherein each node of the dialog tree has only a single parent node and at least 0 child nodes, and wherein as the tree is expanded, at least some of the nodes have at least two child nodes.

7. The method of claim 1, wherein steps a)-e) are repeated a plurality of times.

8. The method of claim 1, further comprising generating a user interface for display to the selected annotator for receiving the annotator'"'"'s annotation.

9. The method of claim 8, wherein the user interface provides a goal of the scenario and a current state of the dialog.

10. The method of claim 1, further comprising conducting a dialog based on the dialog policy, wherein the dialog policy is used to generate a dialog act of the virtual agent in response to the user'"'"'s utterance.

11. The method of claim 1, further comprising outputting at least one of:

the corpus of dialogs; and

a dialog policy generated based on the corpus of dialogs.

12. A computer program product comprising a non-transitory medium storing instructions, which when executed by a computer processor, perform the method of claim 1.

13. A system comprising memory which stores instructions for performing the method of claim 1 and a processor in communication with the memory for executing the instructions.

14. In combination, a system for generating dialogs for learning a dialog policy and a computer-implemented dialog system, the system for generating dialogs comprising:

memory which stores a dialog tree for each of a plurality of scenarios, wherein paths through the tree correspond to respective dialogs and nodes of the tree each represent a turn of a dialog, whereby some of the nodes correspond to user annotations and others of the nodes correspond to agent annotations;

a tree update component for updating the dialog trees based on annotations of annotators in a pool of annotators serving as virtual agents and users;

a reactive tree expansion component which progressively expands the dialog trees by repeated selection of a next node to be annotated by one of the annotators in the pool, the next node being selected based on a respective computed measure of uncertainty for each of the current nodes in one of the dialog trees, whereby when the next node corresponds to a user annotation, a text annotation is provided by the selected annotator for the next node, and when the next node corresponds to an agent annotation, a dialog act is selected by the selected annotator for the node;

a dialog corpus generator which generates a corpus of dialogs from the expanded dialog trees;

a dialog policy learning component which learns a dialog policy based on the corpus of dialogs, the learning of the dialog policy including learning a classifier model which predicts a next action for a current state of a dialog; and

a processor which implements the tree update component, reactive tree expansion component, and dialog corpus generator;

the dialog system being configured for conducting a dialog between a virtual agent and a user, in which the learned dialog policy predicts, based on a state of the dialog, a next action to perform, the action being converted, by the dialog system, to a next utterance of the virtual agent.

15. The method of claim 14, wherein the computing of the measure of uncertainty comprises computing a ranking error according to:

16. The system of claim 14, further comprising at least one of:

a dialog policy refinement component which refines the learned dialog policy; and

an output component which outputs at least one of;

the corpus of dialogs; and

a dialog policy generated based on the corpus of dialogs.

17. In a dialog system which conducts dialogs between a human user of the dialog system and a virtual agent, a dialog policy learned by a method comprising:

storing a respective dialog tree in memory for each of a plurality of scenarios, wherein paths through the tree correspond to respective dialogs and nodes of the tree each represent a turn of a dialog, some of the nodes corresponding to user annotations and others of the nodes corresponding to agent annotations;

progressively expanding the dialog trees by repeated selection of a next node to be annotated by one of a pool of annotators, the next node being selected based on a respective computed measure of uncertainty for each of the nodes currently in the dialog trees, and updating the dialog trees based on the annotation of the one annotators in the pool of annotators, whereby when the next node corresponds to a user annotation, the selected annotator is requested to provide a text annotation for the next node, and when the next node corresponds to an agent annotation, the selected annotator is requested to select a dialog act for the node;

generating a corpus of dialogs from the expanded dialog trees; and

learning a dialog policy based on the corpus;

the dialog system being configured for conducting a dialog between a virtual agent and a user, in which the learned dialog policy predicts, based on a state of the dialog, a next action to perform, the action being converted, by the dialog system, to a next utterance of the virtual agent.

18. A method for learning a dialog policy for a dialog system comprising:

for each of at least one scenario, in which annotators in a pool of annotators serve as both virtual agents and users, generating a respective dialog tree in which each path through the tree corresponds to a dialog and nodes of the tree correspond to turn of a dialog, the generation comprising;

a) computing a measure of uncertainty for nodes in the dialog tree,

b) identifying a next node to be annotated, based on the measure of uncertainty,

c) selecting an annotator from the pool to provide an annotation for the next node,

d) receiving an annotation from the selected annotator for the next node, wherein when the next node corresponds to a user annotation, the received annotation is a text annotation for the next node, and when the next node corresponds to an agent annotation, the received annotation is a dialog act for the next node,

e) generating a new node of the dialog tree based on the received annotation, and

f) repeating a)-e) a plurality of times with different annotators selected from the pool;

generating a corpus of dialogs from the dialog tree; and

based on the corpus of dialogs, learning a classifier model of a dialog policy that predicts a next action for the dialog system, based on a state of a dialog; and

incorporating the learned dialog policy into a dialog system for conducting a dialog between a virtual agent and a user, in which the learned dialog policy predicts, based on a state of the dialog, a next action to perform, the action being converted, by the dialog system, to a next utterance of the virtual agent.