Microsoft Technology Licensing, LLC

Curai, Inc.

Facebook, Inc.

Meta Platforms, Inc.

Anitha Kannan

Simon John Baker

Rakesh Agrawal

Qifa Ke

Ariel Fuxman

Krishnaram Kenthapadi

Sreenivas Gollapudi

Krishnan Ramnath

Dahua Lin

Neil D Hunt

Michael Gamon

Neal Khosla

Patrick Pantel

John Christopher Shafer

Inmar-Ella Givoni

Nikhil Rasiwasia

Xavier Amatriain

Manish Chablani

Thomas Lin

Geoffrey Tso

Murali Ravuri

Vitor Rodrigues

Vignesh Venkataraman

Partha Pratim Talukdar

Matthew Uyttendaele

John Michael Winn

Deepayan Chakrabarti

Carsten Rother

Samuel Ieong

Aditya Pal

Karthik Subbian

David Sontag

Yuandong Tian

Yann Nicolas Dauphin

Juliet Anne Bernstein

Andrew Brian Goldberg

Varun Nair

Elliot Schumacher

Viraj Uday Prabhu

Albert Yu Sun

Marios Kokkodis

Nathaniel Dion Parrish

In one aspect, a method of guarding an automated software includes generating, by a first language model, a training set of rule-violating data, generating, by the first language model, a training set of contrastive examples by altering the rule-violating data into non-violating data, training a guardrail machine learning model using the generated training sets, generating, with an automated software, an output based on a user input, monitoring with the trained guardrail machine learn model whether the generated output violates a rule, and preventing the automated software from transmitting to the user the generated output that violates a rule.

Guardrail machine learning model for automated software

Techniques for responding to a healthcare inquiry from a user are disclosed. In one particular embodiment, the techniques may be realized as a method for responding to a healthcare inquiry from a user, according to a set of instructions stored on a memory of a computing device and executed by a processor of the computing device, the method comprising the steps of: classifying an intent of the user based on the healthcare inquiry; instantiating a conversational engine based on the intent; eliciting, by the conversational engine, information from the user; and presenting one or more medical recommendations to the user based at least in part on the information.

Ensemble machine learning systems and methods

Images stored in an information repository are prepared for browsing. For each image in the repository, text in the repository is mined to extract snippets of text about the image which are semantically relevant to the image, and for each of these snippets of text, keyterms are detected in the snippet of text which represent either concepts that are related to the image or entities that are related to the image, and the snippet of text and keyterms are associated with the image. Each keyterm that is associated with each image in the repository is hyperlinked to each other image in the repository that has this keyterm associated therewith. A graphical user interface allows a user to browse the images in the repository by using their associated snippets of text and hyperlinked keyterms.

Browsing images via mined hyperlinked text snippets

Systems, methods, and non-transitory computer-readable media can train a sequence model to output respective captions, or portions of captions, for content items. A determination can be made that a user of the social networking system is posting a content item for publication through a social networking system. A set of captions, or portions of captions, can be determined for the content item being posted based at least in part on the sequence model. The set of captions, or portions of captions, can be provided as suggestions to the user for use in a caption describing the content item being posted.

Suggesting captions for content

In one embodiment, a system may access a graph data structure that includes nodes and connections between the nodes. Each node may be associated with a user; each connection between two nodes may represent a relationship between the associated users; and each node may be either labeled or unlabeled with respect to a label type. For each labeled node, a label of the label type of that labeled node may be propagated to other nodes through the connections. For each node, the system may store a label distribution information associated with the label type based on the propagated labels reaching the node. The system may train a machine-learning model using the labels and the label distribution information of a set of the labeled nodes. A predicted label for each unlabeled node may be generated using the model and the label distribution information of the unlabeled node.

Semi-supervised learning via deep label propagation

Techniques for responding to a healthcare inquiry from a user are disclosed. In one particular embodiment, the techniques may be realized as a method for responding to a healthcare inquiry from a user, according to a set of instructions stored on a memory of a computing device and executed by a processor of the computing device, the method comprising the steps of: classifying an intent of the user based on the healthcare inquiry; identifying the intent as a need to obtain a dermatological diagnosis or treatment; soliciting, from the user and via the computing device, an image of the user's skin area to be diagnosed or treated; instantiating an image classification system to identify a dermatological disease from the image; and presenting one or more medical recommendations to the user based on the identified dermatological disease.

Systems and methods for responding to healthcare inquiries

In one embodiment, a web service engine servermay predict a successive action by a user based on an entity reference. The web service engine servermay identify an entity referencein a data transmission caused by a user. The web service engine servermay determine from the data transmission a user intention towards the entity referenceusing an intention model based on a transmission log. The web service engine servermay predict a related successive web action optionfor the entity referencebased on the user intention.

Modeling actions for entity-centric search

The presentation style of a video is identified. A set of features that represents the video is computed. A pre-learned video presentation style classifier is then used to weight each of the features in the set of features and determine a presentation style that is predominately employed in the video based on the weighting of the features.

Identifying presentation styles of educational videos

A document such as a book or textbook includes multiple sections such as chapters. Concept phrases are determined for each of the sections based on the text of each section. A set of content items such as videos is received, and each content item is associated with one or more queries that were submitted by users who were provided the content item in a set of search results. These queries are processed to determine concept phrases that are associated with the content items. The content items and their associated concept phrases are compared with the concept phrases associated with the sections to determine, for some or all of the content items, a minimum subset of the sections whose associated concept phrases cover most of the concept phrases that are associated with the content item. The content items are inserted or linked with the sections in their corresponding minimum subsets.

Associating content items with document sections

While a document, such as an e-book, is read by a user on a computing device such as an e-reader, concept phrases are extracted from the document. The extracted concept phrases may be words or phrases that match known concept phrases such as headings. Based on a universal concept phrase graph that includes nodes for each known concept phrase, core concept phrases are determined for the document. These core concept phrases are associated with nodes of the universal concept phrase graph that are located within a predetermined distance of nodes that represent the concept phrases extracted from the document. Each core concept phrase is combined with one or more of the concept phrases to generate multiple queries. These queries are submitted to search engines, and indicators of documents from the corresponding search results are presented to the user with the original document that is being read.

Identifying similar documents using graphs

In one embodiment, a web service engine servermay predict a successive action by a user based on an entity reference. The web service engine serveridentifies an entity referencein a data transmission caused by a user. The web service engine serverdetermines from the data transmission a user intention towards the entity referenceusing an intention model based on a transmission log. The web service engine serverpredicts a related successive web action optionfor the entity referencebased on the user intention.

Systems, methods, and computer storage media are provided for generating rich navigational study aids for electronic books. For a particular section of interest in a document, one or more related sections for providing additional context to the particular section are determined. The related sections are ranked based on a score indicating significance to the particular section. Based on a user's information processing preference, a set of ranked navigational links to each related section is presented to the user for additional context related to the particular section.

Navigational aid for electronic books and documents

Search history data such as browse trails are collected over time. The browse trails, including associated queries and domains, are processed to identify free tokens of the queries that are also modifiers. Attribute value pairs of a structured data source that correspond to the modifiers are determined based on the search history data and a frequency of the attribute value pairs in the structured data source. When a subsequent query is received, modifiers in the query are identified and replaced with the determined combinations of attribute value pairs that correspond to the modifiers in a structured query that is generated from the received query. The structured query is used to identify items and/or services in the structured data source that are responsive to the received query.

Reformulating query terms in structured search

Growing community of inventors

Mountain View, CA, United States of America

Anitha Kannan