Smart replies using an on-device model

• Main Authors: Thomas Matthew Rudick, Nathan Dickerson Beach, Mirko Ranieri, John Patrick McGregor Jr, Sujith Ravi
• Format: Patent
• Language: English
• Published: 06.06.2018
• Subjects: CALCULATING; COMPUTER SYSTEMS BASED ON SPECIFIC COMPUTATIONAL MODELS; COMPUTING; COUNTING; DATA PROCESSING SYSTEMS OR METHODS, SPECIALLY ADAPTED FORADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL, SUPERVISORYOR FORECASTING PURPOSES; ELECTRIC DIGITAL DATA PROCESSING; PHYSICS; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE,COMMERCIAL, FINANCIAL, MANAGERIAL, SUPERVISORY OR FORECASTINGPURPOSES, NOT OTHERWISE PROVIDED FOR

A communication 22 sent from an external computing device is received by a computing device 2 such as a smartphone, laptop, tablet or smart watch. A machine-trained model on the computing device determines candidate responses 28A-C based on the communication. A user selects 29 one of the candidate responses which is then sent to the external computing device. Determining candidate responses may comprise using a random projection function to project the communication into a hash signature, determining a projected node associated with the hash signature, and determining candidate responses from a ranked list of predicted responses. The on-device machine-trained model may be trained by semi-supervised machine learning at an external computer system. The candidate responses may be personalised based on a users communication history, and may include a previous response sent by the user that belongs to the same semantic cluster as a predicted response. The method may be used to suggest candidate responses to communications such as SMS text messages, Internet messages or instant messages in a chat application.