Apple Inc.

Jerome Rene Bellegarda

Jannes G Dolfing

Kim E A Silverman

Giulia Maria Pagallo

Ryan S Dixon

Devang Kalidas Naik

Karl Martin Groethe

Kim Silverman

Jannes G A Dolfing

Lawrence Yuan Yang

Brent D Ramerth

Xin Ru Wang

Alexandre F Carlhian

Karan Misra

Mathieu Jean Martel

Stephen Olivier Lemay

Patrick L Coffman

Christopher Patrick Foss

Richard R Dellinger

Gaurav Kapoor

Cyrus Daniel Irani

Daniel C Gross

Michael R Siracusa

Thomas Deniau

Sofiane Toudji

Alexandre R Moha

Alessandro Francesco Sabatelli

Thomas R Gruber

Harry Joseph Saddler

Douglas R Davidson

Scott James Forstall

Tiffany S Jon

Kevin Tiene

Justin N Wood

Roberto Garcia

Yen-Lu Chow

Gregory N Christie

Joshua H Shaffer

Bertrand Serlet

Wayne Loofbourrow

Steve S Ko

Paul D Marcos

Randal J Marsden

Bronwyn Anne Jones

Xin Wang

Ueli Meier

Mei-Qun Xia

Colin C Morris

Jun Hatori

Jason J Gauci

Karl Christian Kohlschuetter

Hafid J Arras

Aria D Haghighi

Kevin D Clark

Bronwyn A Jones

Bishal Barman

Bryce H Nyeggen

Jesper S Andersen

Kevin Bartlett Aitken

Linden B Siahaan

Jennifer Lauren Moore

Matthias Neeracher

Sibel Yaman

Jesper A Andersen

John William Scalo

Giovanni Donelli

Vivek Kumar Rangarajan Sridhar

Sabrine Rekik

Kevin Lenzo

Bruce Arthur

Alkeshkumar M Patel

Dominic L Howell

John W Butzberger

Justin S Hogg

Antoine Vinel

Andreas Eminidis

Akash Mehra

Markus Berget

Hafid Arras

Richard Harry Starfield

Jannes Ga Dolfing

Kim Ea Silverman

Jannes GA Dolfing

Alexandre Carlhian

Mathieu J Martel

Guilia M Pagallo

The present disclosure generally relates to updating a language model based on user feedback. Based on a user text input, a language model predicts a set of tokens and an action that will be taken by the user in response to the predicted set of tokens. If the predicted action does not match a detected actual user action, the language model is updated to reflect the user feedback by modifying an output token probability distribution based on the actual user action and updating the language model to converge with a target language model using the modified output token probability distribution.

Injection of user feedback into language model adaptation

At an electronic device with a touch-sensitive display, display a message transcript, where the message transcript includes at least one message from at least a first user. Determine, based at least in-part on the at least one message, a plurality of suggested one or more characters. Display, on the touch-sensitive display, the plurality of suggested one or more characters. Detect an input on the touch-sensitive display. Determine whether the input represents user selection of one of the plurality of suggested one or more characters. Display the selected one of the plurality of suggested one or more characters in the message transcript. Send the selected one or more characters to the first user.

Canned answers in messages

Systems and processes for prediction using generative adversarial network and distillation technology are provided. For example, an input is received at a first portion of a language model. A first output distribution is obtained, based on the input, from the first portion of the language model. Using a first training model, the language model is adjusted based on the first output distribution. The first output distribution is received at a second portion of the language model. A first representation of the input is obtained, based on the first output distribution, from the second portion of the language model. The language model is adjusted, using a second training model, based on the first representation of the input. Using the adjusted language model, an output is provided based on a received user input.

Model compression using cycle generative adversarial network knowledge distillation

Systems and methods for analysis and validation of language models trained using data that is unavailable or inaccessible are provided. One example method includes, at an electronic device with one or more processors and memory, obtaining a first set of data corresponding to one or more tokens predicted based on one or more previous tokens. The method determines a probability that the first set of data corresponds to a prediction generated by a first language model trained using a user privacy preserving training process. In accordance with a determination that the probability is within a predetermined range, the method determines that the one or more tokens correspond to a prediction associated with the user privacy preserving training process and outputs a predicted token sequence including the one or more tokens and the one or more previous tokens.

Analysis and validation of language models

Methods, systems, and computer-readable media related to a technique for providing handwriting input functionality on a user device. A handwriting recognition module is trained to have a repertoire comprising multiple non-overlapping scripts and capable of recognizing tens of thousands of characters using a single handwriting recognition model. The handwriting input module provides real-time, stroke-order and stroke-direction independent handwriting recognition for multi-character handwriting input. In particular, real-time, stroke-order and stroke-direction independent handwriting recognition is provided for multi-character, or sentence level Chinese handwriting recognition. User interfaces for providing the handwriting input functionality are also disclosed.

Managing real-time handwriting recognition

Systems and processes for word prediction using multiple contexts are provided. For example, a plurality of words are received. A first word context including a first plurality of received words, and a second word context corresponding to the first plurality of received words and a second plurality of received words, are obtained. A first current word probability is determined based on a first language model using the first word context. A second current word probability is determined based on a second language model using the second word context. A third current word probability is determined based on the second language model using the first word context. A fourth current word probability is determined based on the first current word probability, the second current word probability, and the third current word probability. An output is provided, to a user, including a current word prediction based on the fourth current word probability.

Word prediction with multiple overlapping contexts

Systems and processes are disclosed for handling a multi-part voice command for a virtual assistant. Speech input can be received from a user that includes multiple actionable commands within a single utterance. A text string can be generated from the speech input using a speech transcription process. The text string can be parsed into multiple candidate substrings based on domain keywords, imperative verbs, predetermined substring lengths, or the like. For each candidate substring, a probability can be determined indicating whether the candidate substring corresponds to an actionable command. Such probabilities can be determined based on semantic coherence, similarity to user request templates, querying services to determine manageability, or the like. If the probabilities exceed a threshold, the user intent of each substring can be determined, processes associated with the user intents can be executed, and an acknowledgment can be provided to the user.

Multi-command single utterance input method

Systems and processes for operating an intelligent automated assistant are provided. In one example process a set of words including a grammatical error is received. The process can generate, using a neural network based on the set of words including the grammatical error and a reference set of words, a transformed set of words and further determine, based on the set of words including the grammatical error and the reference set of words, a reconstructed reference set of words. The process can also determine, based on a comparison of the transformed set of words and the reconstructed reference set of words, whether the transformed set of words is grammatically correct and provide an indication of whether the transformed set of words is grammatically correct to the neural network.

Automatic grammar detection and correction

Systems and processes for modifying word predictions are provided. In one example, a user input is received including one or more words. A prediction of a word sequence corresponding to one or more words is obtained, and context information associated with the word sequence is obtained. In accordance with a determination, based on the context information, that the prediction of the word sequence corresponds to a predetermined semantic reference, the prediction of the word sequence is modified, and an output is provided corresponding to the modified prediction of the word sequence. In accordance with a determination, based on the context information, that the prediction of the word sequence does not correspond to a predetermined semantic reference, an output is provided corresponding to the prediction of the word sequence.

Sanitizing word predictions

An example process for recognizing handwritten input includes obtaining input data representing handwritten input, where the handwritten input is associated with a first dimension and a second dimension relative to the handwritten input; sampling the input data to obtain a plurality of coordinates representing the handwritten input; determining, based on the plurality of coordinates, a sequence of vectors representing a respective plurality of portions of the handwritten input, where: each portion of the respective plurality of portions is associated with a respective height and width corresponding respectively to the first and second dimensions, the respective height being greater than the respective width; and consecutive vectors of the sequence of vectors represent respective consecutive portions of the handwritten input; generating, using a handwriting recognition model, based on the sequence of vectors, one or more characters for the handwritten input; and causing the one or more characters to be displayed.

Spatial and temporal sequence-to-sequence modeling for handwriting recognition

An example process includes: obtaining input token(s); determining, using a joint prediction model, based on the input token(s): a first predicted token following the input token(s) and a second predicted token following the first predicted token; and a first user action to be performed on the first predicted token, where determining the first user action includes: determining a first reward value for performing the first user action based on a first current reward value for performing the first user action and a second reward value for performing a second user action on the second predicted token; outputting the first predicted token; detecting a user action performed on the first predicted token; and in accordance with a determination that the detected user action does not match the first user action: causing parameters of the joint prediction model to be updated, the parameters being configured to determine the first user action.

Incorporating user feedback into text prediction models via joint reward planning

Systems and processes for operating an intelligent automated assistant are provided. In one example process, one or more input words can be received. The process can extract, based on the one or more input words, seed data for unsupervised training of a first learning network. Training data that includes a collection of words having typographical errors for the first learning network can be obtained. The process can determine, using the first learning network and based on the seed data and the training data, one or more output words having a probability distribution corresponding to a probability distribution of the training data. The one or more output words can include typographical errors. The process can generate, based on the determined one or more output words, a data set for supervised training of a second learning network. The second learning network can provide one or more typographical error suggestions.

Neural typographical error modeling via generative adversarial networks

Systems and methods for proactively identifying and surfacing relevant content on an electronic device with a touch-sensitive display are disclosed herein. In one aspect, a method includes presenting content in a first application. At least a portion of the content is presented without requiring input from a user. The method further includes receiving a request to open a second application. In response to receiving the request, the second application is presented with an input-receiving field. Before receiving any user input at the input-receiving field, a selectable user interface object is displayed with an indication that the portion of the content was viewed in the first application, allowing the user to paste at least the portion of the content into the input-receiving field. In response to detecting a selection of the selectable user interface object, the portion of the content is pasted into the input-receiving field.

Systems and methods for proactively identifying and surfacing relevant content on an electronic device with a touch-sensitive display

Growing community of inventors

Saratoga, CA, United States of America

Jerome Rene Bellegarda