Multi-user, multi-domain dialog system

• US 10,181,322 B2
• Filed: 12/20/2013
• Issued: 01/15/2019
• Est. Priority Date: 12/20/2013
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Inventors: Hakkani-Tur, Dilek, Tur, Gokhan, Heck, Larry Paul, Wang, Dong
• Primary Examiner(s): Shin, Seong Ah A
• Application Number:
  US14/137,566
• Publication Number:
  US 20150179168A1
• Time in Patent Office: 1,852 Days
• Field of Search: 704 9, 704235, 704275
• CPC Class Codes: G10L 15/22 Procedures used during a speech recognition process, e.g. man-machine dialogue
• CPC Class Codes: G10L 2015/228 of application context

Abstract

A dialog system for use in a multi-user, multi-domain environment. The dialog system understands user requests when multiple users are interacting with each other as well as the dialog system. The dialog system uses multi-human conversational context to improve domain detection. Using interactions between multiple users allows the dialog system to better interpret machine directed conversational inputs in multi-user conversational systems. The dialog system employs topic segmentation to chunk conversations for determining context boundaries. Using general topic segmentation methods, as well as the specific domain detector trained with conversational inputs collected by a single user system, allows the dialog system to better determine the relevant context. The use of conversational context helps reduce the domain detection error rate, especially in certain domains, and allows for better interactions with users when the machine addressed turns are not recognized or are ambiguous.

Claims

1. A computer-implemented method for improving understanding of a computer-addressed conversational input in a multi-user conversation by a human-computer interaction system, the method comprising:

receiving input from on an electronic device from the multi-user conversation;

translating the input into conversational inputs using an input decoder, the conversational inputs including the computer-addressed conversational input and conversational inputs occurring prior to the computer-addressed conversational input, the computer-addressed conversational input instructing a computer to perform an action;

parsing, using natural language processing, conversational inputs into topic segments;

extracting, using an engine trained based on a statistical model, a conversational context of the conversational inputs within at least one of the topic segments, wherein the conversational context for the computer-addressed conversational input is extracted based on at least contextual features obtained from the conversational inputs occurring prior to the computer-addressed conversational input,

wherein the conversational inputs occurring prior to the computer-addressed conversational input include at least one other computer-addressed conversational input and at least one human-addressed conversational input, and the contextual features include (a) a domain of a majority of the conversational inputs occurring prior to the computer-addressed conversational input and (b) a distribution score associated with at least one of the conversational inputs occurring prior to the computer-addressed conversational input.

2. The method of claim 1, further comprising detecting at least one of a slot or an intent of the computer-addressed conversational input.

3. The method of claim 1 wherein the extracting determines a domain of the computer-addressed conversational input.

4. The method of claim 1 wherein the parsing comprises:

collecting the conversational inputs into blocks using a selected window size;

determining a domain distribution score for each block;

evaluating similarity between the blocks using the domain distribution scores;

creating a topic boundary between two blocks when the similarity between the two blocks falls outside of a threshold.

5. The method of claim 4 further comprising calculating an average of domain distribution scores of the conversational inputs contained within each block.

6. The method of claim 4 further comprising:

calculating similarity of the domain distribution scores of two blocks;

comparing the similarity to the threshold.

7. The method of claim 1 further comprising:

collecting a selected number of conversational inputs into a feature extraction group;

selecting a label associated with the conversational inputs in the feature extraction group; and

establishing a feature vector from at least one topic distribution score.

8. The method of claim 7 further comprising selecting the topic distribution score of a prior conversational input within the feature extraction group that is closest in proximity to the computer-addressed conversational input and has a same label as the selected label.

9. The method of claim 7 wherein establishing a feature vector comprises calculating an average topic distribution score for a prior conversational input in the feature extraction group.

10. The method of claim 7 further comprising skipping dialog commands, interjections, and conversational inputs with low domain confidence scores.

11. The method of claim 7 further comprising skipping prior conversational inputs that do not contain nouns or noun phrases.

12. The method of claim 7 further comprising selecting a domain of the computer-addressed conversational input based on the selected label, the feature vector, and a contiguous sequence of n words from the computer-addressed conversational input.

13. A computer readable article of manufacture device containing computer executable instructions which, when executed by a computer, cause the computer to perform a method for detecting a domain of a computer-addressed conversational input in a human-computer interaction system, the method comprising:

receiving input on an electronic device from the multi-user conversation;

translating the input into conversational inputs using an input decoder, the conversational inputs including the computer-addressed conversational input and conversational inputs occurring prior to the computer-addressed conversational input, the computer-addressed conversational input instructing a computer to perform an action;

parsing, using natural language processing, conversational inputs into topic segments;

extracting, using an engine trained based on a statistical model, a conversational context of the conversational inputs within at least one of the topic segments, wherein the conversational context for the computer-addressed conversational input is extracted based on at least contextual features obtained from the conversational inputs occurring prior to the computer-addressed conversational input,

wherein the conversational inputs occurring prior to the computer-addressed conversational input include at least one other computer-addressed conversational input and at least one human-addressed conversational input, and the contextual features include (a) a domain of a majority of the conversational inputs occurring prior to the computer-addressed conversational input and (b) a distribution score associated with at least one of the conversational inputs occurring prior to the computer-addressed conversational input.

14. The computer readable device of claim 13, further comprising detecting at least one of a slot or an intent of the computer-addressed conversational input.

15. The computer readable device of claim 13, wherein the extracting determines a domain of the computer-addressed conversational input.

16. The computer readable device of claim 13, wherein the parsing includes:

collecting the conversational inputs into blocks using a selected window size;

determining a domain distribution score for each block;

evaluating similarity between the blocks using the domain distribution scores;

creating a topic boundary between two blocks when the similarity between the two blocks falls outside of a threshold.

17. A computer system comprising a processor electrically connected to computer readable storage media, the computer-readable storage media storing instructions that, when executed, perform a computer-implemented method for detecting a domain of a computer-addressed conversational input in a human-computer interaction system, the method comprising:

receiving input on an electronic device from the multi-user conversation;

translating the input into conversational inputs using an input decoder, the conversational inputs including the computer-addressed conversational input and conversational inputs occurring prior to the computer-addressed conversational input, the computer-addressed conversational input instructing a computer to perform an action;

parsing, using natural language processing, conversational inputs into topic segments;

extracting, using an engine trained based on a statistical model, a conversational context of the conversational inputs within at least one of the topic segments, wherein the conversational context for the computer-addressed conversational input is extracted based on at least contextual features obtained from the conversational inputs occurring prior to the computer-addressed conversational input,

wherein the conversational inputs occurring prior to the computer-addressed conversational input include at least one other computer-addressed conversational input and at least one human-addressed conversational input, and the contextual features include (a) a domain of a majority of the conversational inputs occurring prior to the computer-addressed conversational input and (b) a distribution score associated with at least one of the conversational inputs occurring prior to the computer-addressed conversational input.

18. The system of claim 17, further comprising detecting at least one of a slot or an intent of the computer-addressed conversational input.

19. The system of claim 17, further comprising:

collecting a selected number of conversational inputs into a feature extraction group;

selecting a label associated with the conversational inputs in the feature extraction group; and

establishing a feature vector from at least one topic distribution.

20. The system of claim 19, further comprising selecting a topic distribution score of a prior conversational input within the feature extraction group that is closest in proximity to the computer-addressed conversational input and has a same label as the selected label.