Techniques for updating a partial dialog state

• US 9,318,109 B2
• Filed: 10/02/2013
• Issued: 04/19/2016
• Est. Priority Date: 10/02/2013
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Inventors: Boies, Daniel, Sarikaya, Ruhi, Rochette, Alexandre, Feizollahi, Zhaleh, Ramesh, Nikhil
• Primary Examiner(s): Pham, Thierry L
• Application Number:
  US14/044,802
• Publication Number:
  US 20150095033A1
• Time in Patent Office: 930 Days
• Field of Search: 704/4, 704/9, 704/14, 704/236, 704/257
• US Class Current: 1/1
• CPC Class Codes: G06F 16/3329 Natural language query formulation or dialogue systems
• CPC Class Codes: G06F 40/30 Semantic analysis
• CPC Class Codes: G10L 15/1815 Semantic context, e.g. disambiguation of the recognition hypotheses based on word meaning

Abstract

Embodiments provide for tracking a partial dialog state as part of managing a dialog state space, but the embodiments are not so limited. A method of an embodiment jointly models partial state update and named entity recognition using a sequence-based classification or other model, wherein recognition of named entities and a partial state update can be performed in a single processing stage at runtime to generate a distribution over partial dialog states. A system of an embodiment is configured to generate a distribution over partial dialog states at runtime in part using a sequence classification decoding or other algorithm to generate one or more partial dialog state hypothesis and/or a confidence score or measure associated with each hypothesis. Other embodiments are included.

Claims

1. A system for updating a partial dialog state, the system configured to:

receive user input at a first turn;

generate a semantic representation based on the user input at the first turn;

generate a partial dialog state based in part on the semantic representation associated with the first turn;

receive user input at a second turn;

generate the semantic representation based on the user input at the second turn;

update the partial dialog state for the second turn using a discriminative model using the semantic representation associated with the first turn and the second turn, wherein the partial dialog state is based on a time ordered sequence of an estimated user goal; and

based in part on the partial dialog state, determining a dialog state as part of tracking a user goal.

2. The system of claim 1, further configured to use a sequence-based classification model at runtime to jointly model slot detection and partial dialog state tracking by detecting slots and updating the partial dialog state.

3. The system of claim 1, further configured to use separate models to detect slots and update the partial dialog state at runtime.

4. The system of claim 1, further configured to:

generate a dialog state using the partial dialog state and knowledge results obtained from one or more knowledge sources; and

execute a policy based on the dialog state.

5. The system of claim 1, further configured to update the partial dialog state before querying any knowledge sources.

6. The system of claim 1, further configured to manage the update of the partial dialog state by integrating the semantic representation of the user input across multiple turns including one or more of the first turn, one or more intermediate turns, and a current turn.

7. The system of claim 6, further configured to generate the partial dialog state to include filled slots that remain relevant for the current turn.

8. The system of claim 1, further configured to perform partial dialog state tracking in part by managing slots to be kept or dropped across turns.

9. The system of claim 1, further configured to generate a distribution over partial dialog states using sequence classification decoding operations to generate one or more hypothesis including a confidence score or probability for each hypothesis.

10. The system of claim 9, wherein the one or more hypothesis provides an estimate of a distribution over a partial dialog state space.

11. A method comprising:

determining a partial dialog state for a turn based in part on user input associated with the turn;

using a discriminative model to update the partial dialog state for a current turn based in part on the partial dialog state associated with the turn, wherein the partial dialog state is a time ordered sequence of an estimated user goal; and

determining a dialog state based in part on the partial dialog state at the current turn as part of tracking a user goal.

12. The method of claim 11, further comprising using a single sequence based classification model at runtime for slot detection and partial dialog state updating.

13. The method of claim 11, further comprising using sequence classification decoding operations to generate a partial dialog state distribution over partial dialog states comprising a finite list of possible states.

14. The method of claim 11, further comprising using the partial dialog state of the current turn to query one or more knowledge sources and determine the dialog state.

15. The method of claim 11, further comprising updating the partial dialog state by determining which filled slots remain relevant in the current turn.

16. The method of claim 11, further comprising updating the partial dialog state using information associated with when each slot was filled and a textual position that slot values originate from.

17. A computer-readable storage device which stores executable instructions that operate to:

determine a semantic representation based on user input at a first turn using natural language processing operations;

determine a partial dialog state for the first turn based in part on the semantic representation associated with the first turn;

determine the semantic representation based on the user input at a second turn; and

update the partial dialog state for the second turn based in part on the semantic representation associated with the first turn and the second turn, wherein the partial dialog state is a time ordered sequence of an estimated user goal; and

based in part on the partial dialog state, determining a dialog state as part of tracking a user goal.

18. The computer-readable storage device of claim 17, wherein the instructions operate further to jointly model slot detection and partial dialog state tracking in part by using a runtime sequence classification technique.

19. The computer-readable storage device of claim 17, wherein the instructions operate further to update the partial dialog state by one of dropping, adding, or modifying a slot.

20. The computer-readable storage device of claim 17, wherein the instructions operate further to use a class to identify words that are not associated with a slot.