Microsoft Technology Licensing, LLC

Michel Galley

William B Dolan

Christopher John Brockett

Christopher Brian Quirk

Jianfeng Gao

Margaret Ann Mitchell

Alessandro Sordoni

Michael Auli

Yangfeng Ji

Zeqiu Wu

Zhang Li

Jian-Yun Nie

Frederik Schaffalitzky

Ibrahim Eden

Kaushik Ramaiah Narayanan

Anthony Aue

Jesse Alexander Freitas

Si-Qing Chen

Warren Anthony Aldred

Weixin Cai

Xinyu He

Yizhe Zhang

Systems and method directed to assistive document generation are described. More specifically, similar documents share large portions of reusable text structures that can be used to generate an initial document thereby saving a user time. To generate the document, an indication to create the document may be received and based on the indication, a plurality of example documents and grounding content may be identified. Example documents may be existing documents that are similar to a target document of the writer. Grounding information may refer to content that is relevant, timely, and accurate when applied to the target document. The plurality of example documents and the grounding content may be received, and a document sketch based on the example documents and the grounding content may be generated and contains a plurality of predicted text sequences based on the example documents and the grounding content.

Automatic document sketching

A controllable grounded response generation framework includes a machine learning model, a grounding interface, and a control interface. The machine learning model is trained to output computer-generated text based on input text. The grounding interface is useable by the machine learning model to access a grounding source including information related to the input text. The control interface is useable by the machine learning model to recognize a control signal. The machine learning model is configured to include information from the grounding source in the computer-generated text and focus the computer-generated text based on the control signal.

Grounded text generation

Systems and methods are directed to generating content that is contextually relevant in a writing style of a user. In example embodiments, a plurality of logical inputs regarding a topic is received in bullet point format. A content generator generates draft content using machine learning (ML) models. The generating comprises identifying a writing style of the user by applying the plurality of logical inputs to a first ML model, determining a context and direction for the draft content using a second ML model, and based on the plurality of logical inputs, the identified writing style, and the context and direction, generating at least one paragraph of draft content in the writing style of the user that follows an outline associated with the bullet point format and comprises a same context and direction as the plurality of logical inputs. The draft content is then presented at a client device.

Automatic content generation

Controllable grounded text generation

Examples are generally directed towards context-sensitive generation of conversational responses. Context-message-response n-tuples are extracted from at least one source of conversational data to generate a set of training context-message-response n-tuples. A response generation engine is trained on the set of training context-message-response n-tuples. The trained response generation engine automatically generates a context-sensitive response based on a user generated input message and conversational context data. A digital assistant utilizes the trained response generation engine to generate context-sensitive, natural language responses that are pertinent to user queries.

Context-sensitive generation of conversational responses

Examples are generally directed towards automatic assessment of machine generated conversational responses. Context-message-response n-tuples are extracted from at least one source of conversational data to generate a set of multi-reference responses. A response in the set of multi-reference responses includes it context-message data pair and rating. The rating indicates a quality of the response relative to the context-message data pair. A response assessment engine generates a metric score for a machine-generated response based on an assessment metric and the set of multi-reference responses. The metric score indicates a quality of the machine-generated conversational response relative to a user-generated message and a context of the user-generated message. A response generation system of a computing device, such as a digital assistant, is optimized and adjusted based on the metric score to improve the accuracy, quality, and relevance of responses output to the user.

Metric for automatic assessment of conversational responses

Automated language translation often involves language translation resources of significant size (e.g., 50-gigabyte phrase tables) and significant computational power exceeding the capabilities of many mobile devices. Remotely accessible servers capable of near-realtime, automated translation may be inaccessible or prohibitively costly while traveling abroad. Presented herein are adaptations of language translation techniques for offline mobile devices involving reducing the size and raising the efficiency of the language modeling resources. A word index may be provided that stores respective string representations of the words of a language, and maps respective words to a location (e.g., address or offset) of respective word representations within the word index. Language translation resources (e.g., phrase tables) may then specify logical relationships using the word index addresses of the involved words, rather than the string equivalents. This technique significantly condenses the language resources and provides faster, bidirectional access to the word representations of the language.

Language processing resources for automated mobile language translation

Growing community of inventors

Seattle, WA, United States of America

Michel Galley