US 11,961,509 B2
Training a user-system dialog in a task-oriented dialog system
Swadheen Kumar Shukla, Seattle, WA (US); Lars Hasso Liden, Seattle, WA (US); Thomas Park, Seattle, WA (US); Matthew David Mazzola, Seattle, WA (US); Shahin Shayandeh, Seattle, WA (US); Jianfeng Gao, Woodinville, WA (US); and Eslam Kamal Abdelreheem, Sammamish, WA (US)
Assigned to Microsoft Technology Licensing, LLC, Redmond, WA (US)
Filed by Microsoft Technology Licensing, LLC, Redmond, WA (US)
Filed on Apr. 3, 2020, as Appl. No. 16/839,308.
Prior Publication US 2021/0312904 A1, Oct. 7, 2021
Int. Cl. G10L 15/00 (2013.01); G06N 3/044 (2023.01); G06N 3/049 (2023.01); G06N 3/08 (2023.01); G10L 15/06 (2013.01); G10L 15/16 (2006.01); G10L 15/22 (2006.01); G10L 25/30 (2013.01)
CPC G10L 15/063 (2013.01) [G06N 3/044 (2023.01); G06N 3/049 (2013.01); G06N 3/08 (2013.01); G10L 15/16 (2013.01); G10L 15/22 (2013.01); G10L 25/30 (2013.01); G10L 2015/0635 (2013.01); G10L 2015/225 (2013.01)] 18 Claims
OG exemplary drawing
 
1. A computer-implemented method for correcting a dialog, the method comprising:
receiving a first dialog graph comprising a plurality of nodes and at least one edge connecting two nodes of the plurality of nodes, wherein the first dialog graph represents a dialog flow including each of the plurality of nodes defining an action associated with the corresponding node and the at least one edge defining a condition linking the two nodes, wherein a first path connects at least a first preceding node to at least a first subsequent node of the plurality of nodes through one or more edges, and wherein a second path connects at least the first preceding node to at least a second subsequent node of the plurality of nodes;
converting the first path of the first dialog graph into a first text-based dialog and the second path of the first dialog graph into a second text-based dialog, wherein the first text-based dialog and the second text-based dialog are in a data format adapted for training a neural network and represent the dialog flow of the first dialog graph;
training the neural network based at least on the first text-based dialog and the second text-based dialog as training data;
receiving a log dialog, wherein the log dialog is generated based on executing the neural network to deploy a dialog, wherein a conversation thread associated with the deployed dialog is included in the log dialog;
identifying an exception in the log dialog;
converting at least a portion of the log dialog into a second dialog graph;
receiving an edit directly to the conversation thread of the deployed dialog in the log dialog via an interactive dialog editor tool to mitigate the exception and to create a corrected dialog;
retraining the neural network based at least on the corrected dialog; and
updating the second dialog graph associated with the log dialog based on the corrected dialog.