Microsoft Technology Licensing, LLC

Facebook, Inc.

Jeffrey William Pasternack

Arpit Dhariwal

David Vickrey

Bing Zhao

Nimesh Madhavan Chakravarthi

Nandeesh Channabasappa Rajashekar

Jia Liu

Giridhar Rajaram

Birjodh S Tiwana

Vita G Markman

Romer E Rosales-Delmoral

Christopher Szeto

Haowen Ning

Max Christian Eulenstein

Adam Leon

Pujita Mathur

Jianfu Chen

Prasoon Mishra

Ryan A Espiritu

Yizhou Pan

Ngoc M Vu

Akshaya Subbaraya Bhat

Justin MacLean Coughlin

Arya Ghatak Choudhury

Austen Norment McDonald

Neil Cruz Pomerleau

Kritarth Anand

Polina Kuznetsova

Howard Mario Narvaez

Chung Yu Wang

Christian Wilhelm Kurt Byza

Kelly K Hosokawa

In an example embodiment, a technique is presented that accesses training data that includes information about items, queries for items, and labels for the combinations of items and queries. The labels may correspond to different events, and there may be multiple different labels for the same combination of item and query. A machine learned model is then trained to learn a function for embedding each item to which a label pertains and a function for embedding each query to which a label pertains. Then, for each item in the training data, the items are embedded using the machine learned model, and the item embeddings for the item are concatenated into a single item embedding. At inference time, a similar concatenation is performed for multiple query embeddings. The concatenated embeddings are then used as input to an approximate k-nearest neighbor search function.

Dynamic utility functions for inference in machine-learned models

Described herein is a technique to facilitate conjunctive filtering for an embedding-based information retrieval system. Each item in a content collection is initially analyzed to identify its property values for some set of properties. For each item, its property values are encoded in the form of a vector, and concatenated with an item embedding for the item to derive an enhanced item embedding. When a query is received, a query embedding is generated. Any filtering parameters received with or as part of the query are encoded in the form of a vector, and the vector is concatenated with the query embedding to derive an enhanced query embedding. The enhanced query embedding is then used in a 'k' approximate nearest neighbor search to identify items relevant to the query and having property values that satisfy the filtering parameters.

Conjunctive filtering with embedding models

Techniques for generating smart replies involving image files are disclosed herein. In some example embodiments, a computer system detects that a first message comprising a first image file has been transmitted from a first computing device of a first user to a second computing device of a second user, and generates a first plurality of smart replies based on a first embedding vector of the first image file, where the first embedding vector of the first image file is based on at least one of first textual metadata of the first image file, first image data of the first image file, and a first set of query text used by a first set of users in a first set of searches that resulted in the first image file being included in a first set of transmitted messages.

Generating smart replies involving image files

Disclosed are systems, methods, and non-transitory computer-readable media for a language proficiency inference system used to determine a user's proficiency in one or more languages. The language proficiency inference system determines both text-based probability scores and profile-based probability scores indicating a probability that a user speaks a language or set of languages. The text-based probability score is based on text associated with the first user, whereas the profile-based probability score is based profile data of the user. The language proficiency inference system determines aggregated probability scores based on the corresponding text-based and profile-based probability scores. For example, the aggregated probability score is the sum of the text and profile-based probability scores. The language proficiency inference system uses the aggregated scores to determine the languages in which the user is proficient.

Language proficiency inference system

Techniques for generating diverse smart replies using a synonym hierarchy are disclosed herein. A computer system may detect that a first set of one or more messages having first content has been transmitted from a first computing device of a first user to a second computing device of a second user, determine a plurality of candidate replies based on the first content of the first set of one or more messages, and then select a plurality of smart replies from the plurality of candidate replies using a hierarchical graph data structure and at least one diversity rule. The selecting of the plurality of smart replies comprises omitting at least one of the plurality of candidate replies from selection based on the at least one diversity rule, which limits a number of the plurality of smart replies that have a common parent node.

Generating diverse smart replies using synonym hierarchy

Methods are presented for classifying responses to communications. One method includes presenting, in a first user interface (UI), a message from a sender to a recipient requesting a response. The first UI includes response options of a first classification, a second classification, and an entry field for entering a free text message. Another operation is for receiving a recipient response with a selection of: (a) the first or the second button and (b) the free text message entered in the text entry field. The method further includes generating a featurized recipient response based on the free text message and the selection. An acceptance machine-learning (ML) model calculates a classification value of the recipient response from a set of possible classification values based on the featurized recipient response. The method further includes presenting, in a second UI for the sender, an indicator for the message and the calculated classification value.

Communication response analyzer

Methods, systems, and computer programs are presented for estimating the language used in a user communication. One method includes an operation for utilizing counters to track use of languages by a user of an online service, the counters being updated based on interactions of the user in the online service. Further, the method includes operations for detecting a text entered by the user and obtaining, by a language classifier, an initial prediction having probabilities for the languages that the text is in the language. A language distribution prediction is calculated based on the initial prediction and the user counters, where the language distribution prediction comprises a probability, for each language, that the text is in the language. Further, the method includes operations for selecting a language used in the text based on the language distribution prediction and causing presentation on a display of a message in the selected language.

Language profiling service

Techniques for generating smart replies are disclosed herein. A computer system may generate candidate replies to a message from a first device of a first user to a second device of a second user based on content of the message using a first model, and determine synonym replies based on the candidate replies using a hierarchical graph data structure, with the synonym replies including the candidate replies in addition to synonyms of the candidate replies, the hierarchical graph data structure comprising a tree of concepts ranging from root nodes to leaf nodes of synonym replies. The computer system may generate smart replies using a second model based on the synonym replies and corresponding user selection data for each one of the plurality of synonym replies, with the user selection data indicating a number of times the second user has selected the corresponding synonym reply for replying to messages.

Personalization and synonym hierarchy for smart replies

Systems, devices, media, and methods are presented for dynamically sampling training elements while training a plurality of machine learning models. The systems and methods access a plurality of training elements and generate an element buffer including a specified number of positions. The systems and methods generate a random value for each position for the element buffer and map a training element of the plurality of training elements to each position within the element buffer in a first order. The systems and methods generate a second order for the training elements and allocate the training elements to a plurality of machine learning models according to the second order.

Sequential sampling from multisets

Methods, systems, and computer programs are presented for managing counters with automatic obsolescence of outdated values. One method includes operations for initializing a counter at a first time, and receiving, at a second time, a request to add a value to the counter. Further, a time period corresponding to the second time is calculated, as well as a promotion factor based on the time period. The method further includes calculating an incremental value based on the received value and the promotion factor, and adding the incremental value to the counter. Additionally, a prediction or estimate is based on a current value of the counter, such as identifying areas of interest for a user on an online service, identifying languages spoken by the user, strength of connections between users, determining if the user is actively searching for a new job, and so forth.

Counter with obsolescence of outdated values

Applications are presented for selecting response suggestions in a Graphical User Interface (GUI). Responses are identified for a message received in the GUI. Each response has a score and occupies a width of pixels in the GUI, which provides an available width of pixels for presenting the responses. Further, the method includes an operation for identifying sets of the responses, each set having one or more of the suggested responses. For each set, a utility of the set is calculated based on the score of the responses in the set, the width of pixels of the responses in the set, and the available width of pixels. Further, the set with the greatest utility is selected and then the responses of the set are presented in the GUI. Same principles may be applied to organizing screen layouts for applications that scroll vertically, such as user feeds or search results.

Determination of best set of suggested responses

Disclosed are systems, methods, and non-transitory computer-readable media for using a language classification system for generating response messages. A messaging system receives a message transmitted from a first user to a second user, and determines, based on a set of language counter values determined for the second user, a probability value that the second user will respond to the message in a first language and a probability value that the second user will respond to the message in a second language. The language counter values are determined using a text classification model and indicate a number of times that the second user has used the languages in previous messages. Based on the probability values, the messaging system determines that the second user will respond to the message in the first language and causes a set of recommended responses in the first language be presented on the second client device.

Language classification system

Disclosed are systems, methods, and non-transitory computer-readable media for an improved text classification model. A text classification system determines a first embedding for a first set of characters in a text, and a second embedding for a second set of characters in the text. The text classification system applies a first coefficient to the first embedding and a second coefficient to the second embedding, yielding a first weighted embedding and a second weighted embedding. The first coefficient is different than the second coefficient. The text classification system determines a weighted average embedding for the text based on the first weighted embedding, the second weighted embedding, the first coefficient and the second coefficient. The text classification system identifies at least a first classification label and a second classification label for the text based on the weighted average embedding for the text.

Quick text classification model

An online system predicts topics for content items. The online system provides one or more topic labels for a user to apply concurrently while a user is composing a post, in response to requests periodically received from the user's device. A request includes information such as content composed by the user and contextual information. The online system employs machine learning techniques to analyze content composed by a user and contextual information thereby to predict topic labels. Different machine learning models for classifying individual topic labels, identifying relevant topic labels, and/or detecting changes in existing topic predictions are developed. Some machine learning models predict topics for full content and some predict topics for partial content. The online system trains the machine learning models to ensure accurate topic predictions are provided timely. The online system employs various machine learning model training methods such as active training and gradient training.

Growing community of inventors

Belmont, CA, United States of America

Jeffrey William Pasternack