Microsoft Technology Licensing, LLC

Hung-Chih Yang

Xiong Zhang

Danny B Lange

Graham Arthur Wheeler

Bhalchandra Pandit

Jason J Joyce

David M Moore

Cezary Antoni Marcjan

Robert L Rounthwaite

Yi-Min Wang

Johnson T Apacible

Li-wei He

Peter Bailey

Aravind Bala

Shoou-Jiun Wang

Scott Kirk Imig

David Shigeru Taniguchi

Chao Liu

Jinliang Fan

Jayaram Nm Nanduri

Gayathri Ravichandran Geetha

Ming-Yu Fan

Tien Widya Suwandy

In non-limiting examples of the present disclosure, systems, methods and devices for training conversational language models are presented. An embedding library may be generated and maintained. Exemplary target inputs and associated intent types may be received. The target inputs may be encoded into contextual embeddings. The embeddings may be added to the embedding library. When a conversational entity receives a new natural language input, that new input may be encoded into a contextual embedding. The new embedding may be added to the embedding library. A similarity score model may be applied to the new embedding and one or more embeddings for the exemplary target inputs. Similarity scores may be calculated based on the application of the similarity score model. A response may be generated by the conversational entity for an intent type for which a similarity score exceeds a threshold value.

Natural language processing models for conversational computing

Training risk determination models based on a set of labeled data transactions. A first set of labeled data transactions that have been labeled during a review process is accessed. A first risk determination model is trained using the first set of labeled data transactions. A first risk score for data transactions of a set of unlabeled data transactions is determined using the first risk determination model. Data transactions in the set of unlabeled data transactions are newly labeled based on the first risk score. The newly labeled data transactions are added to a second set of labeled data transactions that include the first set of labeled data transactions. A second risk determination model is trained using at least the second set of labeled data transactions. A second risk score is determined for subsequently received data transactions and these data transactions are rejected or approved based on the second risk score.

Using semi-supervised label procreation to train a risk determination model

A system includes a processing device configured to execute instructions including accessing messages sent by a first user and other users. The instructions include identifying a first subset of messages that were sent by the first user and identifying a second subset of messages that are related to a first topic and were sent by the first user. The instructions include determining an authority value based on counts of the first and second subsets. The instructions include identifying a third subset of messages that resulted from the other users forwarding the second subset. The instructions include determining a power value based on a count of users that sent the third subset and computing an impact score for the first user on the first topic based on the authority and power values. The instructions include, based on the impact score, contacting the first user in relation to the first topic.

Computer system for automated assessment at scale of topic-specific social media impact

A system includes a memory storing a dataset representing a community of users of a social networking service, the social networking service providing a platform enabling a user within the community to communicate with at least one of a plurality of other users within the community. The system includes a processor coupled to the memory, the processor configured to determine a ranking of the service users in the dataset based on an initial influence score for at least one of the users. The processor is also configured to determine a contribution to a first user within the community by at least one second user of the plurality of other users within the community, and calculate an influence score for the first user. The ranking is revised based at least in part on the calculated influence score, and information is rendered to a target user based on the revised ranking.

Identifying influential users of a social networking service

A canopy clustering process merges at least one set of multiple single-center canopies together into a merged multi-center canopy. Multi-center canopies, as well as the single-center canopies, can then be used to partition data objects in a dataset. The multi-center canopies allow a canopy assignment condition constraint to be relaxed without risk of leaving any data objects in a dataset outside of all canopies. Approximate distance calculations can be used as similarity metrics to define and merge canopies and to assign data objects to canopies. In one implementation, a distance between a data object and a canopy is represented as the minimum of the distances between the data object and each center of a canopy (whether merged or unmerged), and the distance between two canopies is represented as the minimum of the distances for each pairing of the center(s) in one canopy and the center(s) in the other canopy.

Grouping documents and data objects via multi-center canopy clustering

This patent relates to a system for identifying influential users of a social network platform. The system may access records of messages sent or received by a plurality of users. The system can determine a power of a user to reach other users based at least in part on the messages. The system can determine an authority of the user on a topic. The system can also compute an impact score that reflects both the power and the authority of the user on the topic. Such scores may be used to select users based on impact as part of social trend analysis, marketing or other functions.

Social media impact assessment

A system for identifying influential users of a social network platform. The system may compute a score for each of multiple users. Such a score may be topic-based, leading to a more accurate identification of influential users. Such a topic-based score may indicate authority and/or impact of a user with respect to a topic. The impact may be computed based on authority combined with other factors, such as power of the user. The authority score may be simply computed, in whole or in part, directly from a tweet log without, for example creating a retweet graph. As a result, the scores may be computed, using MapReduce primitives or other constructs that allow the computations to be distributed across multiple parallel processors. Such scores may be used to select users based on impact as part of social trend analysis, marketing or other functions.

Technologies pertaining to computing a metric that is indicative of whether an n-gram in a large corpus of text is a collocation are described herein. The metric is computed in connection with a distributed computing framework, wherein n-grams of varying lengths can be analyzed in a single input data pass, and wherein secondary sorting functionality of the distributed computing framework need not be invoked.

Identifying collocations in a corpus of text in a distributed computing environment

Techniques for identifying influential users of a social networking service are provided. Influential users may be identified via an algorithm in which an influence score is assigned to each user based at least in part on other members of the community users having taken an affirmative step with respect to the user's communications. Iterative processing may be performed, with each user's influence score being determined by contributions from other users, and each contribution being determined by the contributor's influence score as of a prior iteration. A map-reduce framework may be employed, with data representing the community being partitioned into a plurality of discrete shards, a map process corresponding to each shard calculating an influence score for users represented in the shard, and reduce processes ranking users according to influence score across all shards.

Architecture that automatically detects historical search contexts as well as behaviors related to a search query. Machine learning and hand-authored rules are employed to automatically identify search contexts. Historical information likely to be useful in the current context is surfaced. When a user enters a search query or executes another search behavior, past behaviors are exposed which are contextually related to the current behavior. The architecture also provides automatic discovery of historical contexts, features related to the contexts, and training or authoring of a system for classifying behavior into contexts, using some combination of the machine learning and/or hand-authored rules. A runtime system classifies the current user behavior into a context and surfaces contextual information to the user.

Auto-detection of historical search context

Multi-center canopy clustering

Architecture that scales up the non-negative matrix factorization (NMF) technique to a distributed NMF (denoted DNMF) to handle large matrices, for example, on a web scale that can include millions and billions of data points. To analyze web-scale data, DNMF is applied through parallelism on distributed computer clusters, for example, with thousands of machines. In order to maximize the parallelism and data locality, matrices are partitioned in the short dimension. The probabilistic DNMF can employ not only Gaussian and Poisson NMF techniques, but also exponential NMF for modeling web dyadic data (e.g., dwell time of a user on browsed web pages).

Growing community of inventors

Bellevue, WA, United States of America

Hung-Chih Yang