US 10,706,086 B1Granted

Collaborative-filtering based user simulation for dialog systems

AssigneeAmazon Technologies, Inc.

InventorsDevin Didericksen, Jared Kane Kramer, Oleg Rokhlenko, Kevin Michael Small, Li Zhou

Life of this patent

Estimating the term…

Loading patent…

Abstract

Techniques for simulating a user in a conversation are described. A user simulation service and a conversation agent service conduct a dialog. The user simulation service compares a current sequence of stored labels corresponding to statements in the dialog with a plurality of candidate sequences of labels corresponding to statements in a plurality of candidate dialogs to identify a matching sequence of labels. The user simulation sequence identifies a base sequence of labels that includes the matching sequence of labels to identify a label corresponding to an act in the base sequence of labels following the matching sequence of labels. The user simulation service issues the act to the conversation agent service to be used as a simulated user act.

Claims20

Dependent claims

A computer-implemented method comprising:

receiving an agent statement from a conversational agent service that simulates a conversation with a user;
adding the agent statement to a log of a current conversation with the conversational agent service;
classifying the agent statement with a label;
updating a current sequence of labels with the label, wherein each label in the sequence of labels has a corresponding statement in the log;
calculating a first score by comparing the current sequence of labels with a first candidate sequence of labels corresponding to a first sequence of statements in a first candidate conversation, the first candidate sequence of labels preceding a first candidate user statement;
calculating a second score by comparing the current sequence of labels with a second sequence of labels corresponding to a second sequence of statements in a second candidate conversation, the second candidate sequence of labels preceding a second candidate user statement;
and
outputting the first candidate user statement when the first score exceeds the second score.

5 more elements

Dependencies are inferred from the claim text.