Microsoft Technology Licensing, LLC

Oracle International Corporation

Maplebear Inc.

Sap Ag

Omar Rogelio Alonso

Vasileios Kandylas

Serge-Eric Tremblay

Martin Hentschel

Sushma Nagesh Bannur

Kartikay Khandelwal

Meeten Bhavsar

Thomas Chang

Ciya Liao

Joaquin Delgado

Chad Carson

Ramasubramanian Balasubramanian

Ho John Lee

Girija Narlikar

Karen Shiells

Thibault Sellam

Lei Duan

Atul S Kumar

Frances James

Siddhartha Sen

Jake M Hofman

Rukmini Iyer

Shankar Kalyanaraman

Aleksandrs Slivkins

Justin Ormont

Paul Alexander Dow

Ittai Abraham

Thibault Henri Joseph Sellam

Hemant Purohit

Vasilis Kandylas

Travis Alexander Gingerich

An online concierge system generates recipe embeddings for recipes including multiple items and user embeddings for users, with the recipe embeddings and user embeddings in a common latent space. To generate the user embeddings and the recipe embeddings, a model includes separate layers for a user model outputting user embeddings and for a recipe model outputting recipe embeddings. When training the model, a weight matrix generates a predicted dietary preference type for a user embedding and for a recipe embedding and adjusts the user model or the recipe model based on differences between the predicted dietary preference type and a dietary preference type applied to the user embedding and to the recipe embedding. Additionally cross-modal layers generate a predicted user embedding from a recipe embedding and generate a predicted recipe embedding from a user embedding that are used to further refine the user model and the recipe model.

Method, computer program product, and system for training a machine learning model to generate user embeddings and recipe embeddings in a common latent space for recommending one or more recipes to a user

Aspects of the technology described herein detect experts in social networks. Experts may be identified within social networks using social network data. A collection of topics may be compiled based on search engine behavioral data. A combination of the social network data and search engine behavioral data is utilized to score each topic within potential topics for a user. When a topic score is over a predetermined threshold, a user may be classified as an expert for that topic. Expert search interfaces may be generated for use in browsing or searching for experts within a topic.

Expert detection in social networks

Described herein are technologies relating to generating search results responsive to receipt of a query. More specifically, the query is mapped to a topic in response to receipt of a query, and social media accounts that have been labeled as being knowledgeable on the topic are identified. Messages in a message feed of the social media account that are germane to the topic are retrieved, and documents referenced (linked) in the retrieved messages are identified. These documents are positioned in a ranked list based upon the documents being referenced in the messages.

Search engine results for low-frequency queries

Examples are disclosed that relate to grouping query clusters based on intent of a search query. One example provides a system for generating recommended queries related to a search query, the system including a logic subsystem, and memory holding instructions executable by the logic subsystem to receive the search query, execute the search query on an index of pre-computed query clusters to identify a baseline query cluster and one or more neighboring query clusters, form a data structure including queries from within the baseline query cluster and the one or more neighboring query clusters, traverse the data structure to form a hierarchical arrangement of queries, and send the hierarchical arrangement of queries to a computing device from which the search query was received.

Providing query recommendations

Systems and methods are presented for generating a transition graph from check-in data for determining probabilistic transitions of a user from an origin (location) to a destination. Check-in data for a plurality of users is obtained. The check-in data is organized and analyzed to determine a set of transition tuples for each of the plurality of users. These transition tuples are aggregated and used to generate a transition graph comprising a plurality of nodes and edges. Each node corresponds to a location and each edge indicates a transition from an origin node to a destination node and is associated with a likelihood of a user transitioning from the origin node to the destination node.

Generating probabilistic transition data

One or more techniques and/or systems are provided for populating an information interface based upon social media data. For example, users may post, share, and/or discuss various information through social media sources. Accordingly, social media data may be obtained from such social media sources. The social media data may be grouped into sets of social media data based upon temporal information. Within the sets of social media data, social media entries may be clustered into topic clusters (e.g., a royal wedding topic cluster, a plane crash topic cluster, etc.). Event summaries may be generated for respective topic clusters. The event summaries may be used to populate timeslots of an information interface, such as a calendar or timeline, to create annotated timeslots. In this way, the information interface may provide users with an interactive view of events over a time period, such as a year-in-review, based upon social media data.

Social media driven information interface

A technique is described for generating a knowledge graph that links names associated with a first subject matter category (C1) (such as brands) with names associated with a second subject matter category (C2) (such as products). In one implementation, the technique relies on two similarly-constituted processing pipelines, a first processing pipeline for processing the C1 names, and a second processing pipeline for processing the C2 names. Each processing pipeline includes three main stages, including a name-generation stage, a verification stage, and an augmentation stage. The generation stage uses a voting strategy to form an initial set of seed names. The verification stage removes noisy seed names. And the augmentation stage expands each verified name to include related terms. A final edge-forming stage identifies relationships between the expanded C1 names and the expanded C2 names using a voting strategy.

Resource-efficient generation of a knowledge graph

Representative embodiments disclose mechanisms to compile documents into a timeline document that tracks the evolution of a topic over time. Social media documents can be used to identify importance or popularity of linked documents (i.e., documents shared by social media in a post, tweet, etc.). A collection of social media documents is analyzed and used to identify a series of n-grams and a ranked list of linked documents. A subset of the ranked list is selected based upon similarity to the series of n-grams. The subset is then summarized and captured, along with underlying supporting data, into an entry of a timeline document. Related entries in different timeline documents can be linked to create a pivot point that allows a user to jump from one timeline to another. Timeline documents can be made available as part of a search performed by a query system.

Compiling documents into a timeline per event

A user who is traveling through a region may receive information from a device, such as vehicle status (e.g., speed and fuel status); the current environment (e.g., ambient temperature and weather); the user's route (e.g., traffic and accident notification); and/or a content stream (e.g., a radio or podcast). However, such information may be unrelated to the user's focus or motivation for traveling, and may be distracting. The user may request and receive information about the travel topic, but some users may not have available attention and/or the impetus to initiate such requests. Instead, a device may detect a current location and direction of the user; predict the user's destination according to the current location and direction; and retrieve and present to the user one or more content items that are associated with the predicted destination (e.g., breaking news alerts involving an entity or organization located at the user's predicted destination).

Content presentation based on travel patterns

Aspects of the technology described herein generate a comment-summary interface that can help a user find comments of interest to the user. The comment-summary interface allows a user to access comments according to topics instead of scrolling through all of the comments found in a comment section associated with a primary content, such as a news article on a website, a social post, product review, and such. Categorizing the comments into topics makes more efficient use of computer resources by avoiding the need to display comments that are not of interest to the user. The plurality of comments analyzed by aspects of the technology described herein are unstructured comments. Unstructured comments lack specific subject matter categories designated by a user or preselected by a webpage.

Content categorization

One or more techniques and/or systems are provided for identifying relationships between entities and/or for classifying entities into sectors. Social network commentary of users may be evaluated to select comments having co-occurrences of entity identifiers and/or co-occurrences of entity identifiers with sector identifiers. For example, a microblog message 'I love these stores!! $CO(A) $CO(B) #ShoeShopping' may have a co-occurrence of entity identifiers $CO(A) of a Company (A) entity and $CO(B) of a Company (B) entity along with co-occurrence with a #ShoeShopping sector identifier of a Retail Clothing sector. A relationship between the Company (A) entity and the Company (B) entity may be specified based upon the co-occurrence of the entity identifiers. A sector classification of the Retail Clothing sector may be assigned to the Company (A) entity and to the Company (B) entity based upon the co-occurrence of the entity identifiers $CO(A) and $CO(B) with the #ShoeShopping sector identifier.

Entity classification and/or relationship identification

One or more techniques and/or systems are provided for metadata tag evaluation. For example, a metadata tag, associated with content, may be identified (e.g., a hashtag #ML may be used to tag a social network post). A set of characters, within the content, may be evaluated utilizing a probability matrix and the content to identify an expanded metadata tag (e.g., an expanded hashtag 'machine learning'). Descriptive content, such as websites, articles, social network posts, and/or other content associated with the expanded metadata tag, may be retrieved. A description for the metadata tag may be generated based upon the descriptive content (e.g., a definition for machine learning). In this way, the description, related metadata tags, and/or supplemental content may be provided to users having an interest in learning about the metadata tag.

Metadata tag description generation

Systems and methods for extracting and organizing quantitative content with a corresponding topic in regard to a particular, desired event, from a body of user posts is presented. More particularly, nearly ubiquitously a body of user posts includes a substantial amount of highly relevant and interesting information goes largely unprocessed and widely inaccessible. According to the disclosed subject matter, a body/corpus of user posts is filtered according to a desired event such that those user posts relating to a desired event is identified. Additionally, the user posts are also filtered according to whether or not the user posts include a quantitative value. An analysis of the filtered user posts is conducted to extract, for qualifying user posts, a quantitative tuple comprising a quantitative value and a corresponding topic.

Growing community of inventors

Redwood City, CA, United States of America

Omar Rogelio Alonso