Microsoft Technology Licensing, LLC

Yahoo! Inc.

Ahmed Hassan Awadallah

Ryen William White

Patrick Pantel

Michael Gamon

Eric J Horvitz

Madian Khabsa

Yu Su

Miaosen Wang

Imed Zitouni

Ashequl Qadir

Wei W Wang

Yi-Min Wang

Christopher Brian Quirk

Paul Nathan Bennett

Xiao Yang

Wei Dong Wang

Mark James Encarnacion

Subhabrata Mukherjee

Hany Mohamed Hassan Awadalla

Saghar Hosseinisianaki

Yang Song

Susan Theresa Dumais

Nachiappan Nagappan

Rosie Jones

Chetan Bansal

Fernando David Diaz

Aidan C Crook

Thomas Michael Josef Zimmermann

Adish Singla

Frederic H Behr, Jr

Liwei He

Milad Shokouhi

Bahareh Sarrafzadeh

Kyle Mark Williams

Chia-Jung Lee

Bradley Wethington

Chu-Cheng Lin

Dongyeop Kang

Foyzul Hassan

Christopher Huai-Hsien Lin

A method for training a machine learning model with parallel annotations of source instances and while facilitating security of the source instances can be performed by a system that generates a coupled machine learning model from (1) a first machine learning model trained on a first set of training data comprising unannotated natural language and (2) a second machine learning model trained on populated target templates which are populated with a plurality of vocabulary words. Once formed, the coupled machine learning model is configured to transform unannotated natural language into annotated machine-readable text.

Unsupervised method to generate annotations for natural language understanding tasks

Biases in search and retrieval (i.e., situations where searchers seek or are presented with information that significantly deviates from the truth) may be detected by comparison to one or more authoritative sources. Once bias or potential bias is detected, techniques may be applied to indicate and/or compensate for the bias. Such techniques may allow users to more easily assess the veracity of search results, and increase the chances that users will locate accurate answers to their questions.

Compensating for bias in search results

Representative embodiments disclose mechanisms to map natural language input to an application programming interface (API) call. The natural language input is first mapped to an API frame, which is a representation of the API call without any API call formatting. The mapping from natural language input to API frame is performed using a trained sequence to sequence neural model. The sequence to sequence neural model is decomposed into small prediction units called modules. Each module is highly specialized at predicting a pre-defined kind of sequence output. The output of the modules can be displayed in an interactive user interface that allows the user to add, remove, and/or modify the output of the individual modules. The user input can be used as further training data. The API frame is mapped to an API call using a deterministic mapping.

Natural language to API conversion

A computer implemented method includes receiving an exception generated based on programming code, generating exception features from the received exception, the generated exception features being generated based on a set exception features derived from search logs, and executing a machine learning model on the received exception and generated exception features to provide information from the search logs identified as most helpful to resolve the received exception, wherein the machine learning model was trained on training data comprising extracted exceptions and the set of exception features derived from the search logs.

Automated exception featurization and search

Generally discussed herein are devices, systems, and methods for identifying a purpose of a communication. A method can include receiving a communication including communication content and communication context, the communication content a first portion of the communication and the communication context a second, different portion of the communication. The method can include identifying, by a machine learning (ML) model, based on the communication content and the communication context, one or more purposes associated with the communication, the one or more purposes indicating respective actions to be performed by a user that generated or received the communication. The method can include providing data indicating the purpose of the first portion of the content.

Purpose detection in communications using machine learning

A method for applying a trained machine learning model to answer a user query comprises receiving a query text from a user. A previously-trained discriminator is received, the previously-trained discriminator configured to output, for a pair of sentences, a match value indicating a quality of semantic match between the pair of sentences. For each candidate answer text of a plurality of candidate answer texts, the previously-trained discriminator is operated to output a candidate match value for the query text and the candidate answer text based on comparing a first hierarchy of representations of the query text at increasing degrees of semantic compression to a second hierarchy of representations of the candidate answer text at increasing degrees of semantic compression. An answer text is output that is associated with a highest candidate match value among candidate match values.

Multi-scale model for semantic matching

This disclosure describes techniques and architectures that involve a latent activity model for workplace emails. Such a model is based, at least in part, on a concept that communications, such as email at a workplace, are purposeful and organized by activities. An activity is a set of interrelated actions and events around a common goal, involving a particular group of people, set of resources, and time framework, for example. The latent activity model involves a probabilistic inference in graphical models that jointly captures the interplay between latent activities and the email contexts governed by the emails. Such contexts may be email recipients, subject and body of the email, and so on.

Activity modeling in email or other forms of communication

A method includes receiving an email addressed to a recipient user, processing the received email using a reparametrized recurrent neural network model to identify an action based on the received email, and wherein the reparametrized recurrent neural network model has been trained on an email dataset annotated with recipient corresponding actions and reparametrized on unannotated conversation data having structures similar to email data.

AI system to determine actionable intent

Subject matter involves using natural language to Web application program interfaces (API), which map natural language commands into API calls, or API commands. This mapping enables an average user with little or no programming expertise to access Web services that use API calls using natural language. An API schema is accessed and using a specialized grammar, with the help of application programmers, canonical commands associated with the API calls are generated. A hierarchical probabilistic distribution may be applied to a semantic mesh associated with the canonical commands to identify elements of the commands that require labeling. The identified elements may be sent to annotators, for labeling with NL phrases. Labeled elements may be applied to the semantic mesh and probabilities, or weights updated. Labeled elements may be mapped to the canonical commands with machine learning to generate a natural language to API interface. Other embodiments are described and claimed.

Natural language interface to web API

Described herein are technologies pertaining to determining which search engine results page (SERP), from a plurality of SERPs, is preferable to a user. A query is received, and multiple SERPS are retrieved based upon the query, wherein the multiple SERPs are generated independently from one another. Values of features of the query and the multiple SERPs are obtained, and a determination as to which of the SERPs from the multiple SERPs is preferable to the user is made based upon the values of the features. The SERP determined to be preferable over other SERPs in the multiple SERPs is presented to the user.

Identifying preferable results pages from numerous results pages

A method for estimating a quality of semantic match of a first sentence to a second sentence comprises outputting a first hierarchy of representations of the first sentence at increasing degrees of semantic compression, outputting a second hierarchy of representations of the second sentence at increasing degrees of semantic compression, comparing a selected representation in the first hierarchy to each of a plurality of representations in the second hierarchy, comparing a selected representation in the second hierarchy to each of a plurality of representations in the first hierarchy, and outputting a match value indicating a quality of semantic match between the first sentence and the second sentence, the match value based at least on the comparisons.

Growing community of inventors

Redmond, WA, United States of America

Ahmed Hassan Awadallah