Microsoft Technology Licensing, LLC

Robin Abraham

Vijay Mital

Liang Du

Suraj T Poozhiyil

Maurice Diesendruck

Ranjith Narayanan

Darryl Ellis Rubin

Olivier Colle

Sandeep Suresh

Harsh Shrivastava

Nizam Anuar

Henry Hun-Li Reid Pan

Arnaud Christian Flutre

Faisal Khaled Faisal Ilaiwi

Pramod Kumar Sharma

Bao Quoc Nguyen

Xiaofei Zeng

Anthony Wah Lee

J Brandon Smock

Fahimeh Raja

Natalia Larios Delgado

George Edward Busby

Curtis DeSantis

Andrew V Fawcett

Zeeshan Ahmed

Leo Moreno Betthauser

Paul Pangilinan Del Villar

Wenhan Wang

Mingyang Xu

Andy Daniel Martinez

Ziyao Li

Sarah Panda

Jing Tian

Yijian Xiang

Paishun Ting

Mohammad Reza Sarshogh

Jedediah Miller Singer

Henrique Dantas

Sacha Arozarena Valladare

Stuart Richard Long

Dustin James Stewart

Sam Daniel Gammon

Timothy Earnheart

Lipsa Patnaik

Jianzhe Lin

Julia Chen

Manqing Mao

Min Xiao

Nut Limsopatham

Rohith Venkata Pesala

Owen Stephenson Whiting

Yao Li

Sudipto Mukherjee

Ke Jiang

The present disclosure relates to methods and systems for using large language models to support research activities. The methods and systems include a copilot engine that creates input prompts to provide to the large language model to use in generating responses to input messages. The copilot engine infers an intent of the input messages and sends the intent with the input message in the input prompt to the large language model. The large language model generates different types of responses for different intents.

Research activities through conversational user experience using multi-modal large pretrained models

The present disclosure relates to extracting entities from a collection of digital content items based on text from within the digital content items. For example, the present disclosure describes a customizable entity extraction system that utilizes a number of models to extract entities, rank entities, and classify certain entities using a combination of rule-based and machine learning approaches. In one or more embodiments, a customizable entity extraction system applies a set of rules to unstructured text of a collection of digital content items to extract and classify a set of entities in connection with a specific domain of interest.

Extracting and classifying entities from digital content items

Systems and methods are provided for generating a training dataset for training a molecule embedding module using contrastive learning, wherein the definition of similarity is based on molecular scaffold similarity. For example, systems access a molecular dataset and separate the molecular dataset into positive samples and negative samples. Systems then generate a training dataset comprising the positive samples and negative samples. Systems and methods are also provided for using the trained molecule embedding module to generate molecule embeddings and for building an end-to-end machine learning model configured to perform molecular embedding analysis and molecular property prediction, the model comprising the trained molecule embedding module and a property prediction module.

Pre-training molecule embedding GNNs using contrastive learning based on scaffolding

A computer implemented method includes rendering a document page as an image; detecting tables, columns, and other associated table objects within the image via one or more table recognition models that model objects in the image as overlapping bounding boxes; transforming the set of objects into a structured representation of the table; extracting data from the objects into the structured representation; and exporting the table into the desired output format.

Inferring structure information from table images

The present disclosure relates to methods and systems for sharing with a plurality of users a chat session that uses large language models to provide responses for input messages received for the chat session. The methods and systems provide access to the chat session to the users and update the chat session in response to any changes made to the chat session by any of the users. The methods and systems allow the users to resume the chat session at a future time using the chat session history.

Copilot for multi-user, multi-step collaborations

The present disclosure relates to systems, methods, and computer-readable media for utilizing an interactive graphing system to achieve improved dataset exploration utilizing an intelligent workflow and an interactive user interface. More specifically, the interactive graphing system facilitates generating updated network graphs that include inferred user influences based on implicit user action. Indeed, the interactive graphing system can automatically generate and present a user with an updated network graph that includes added, removed, or subsetted elements and relationships that are otherwise hidden from a user. Additionally, the interactive graphing system facilitates network graph exploration and processing of customized combined network graphs that join otherwise separate network graphs.

Generating real-time inferred network graphs

This disclosure describes a machine learning system that includes a contrastive learning based two-tower model for retrieval of relevant chemical reaction procedures given a query chemical reaction. The two-tower model uses attention-based transformers and neural networks to convert tokenized representations of chemical reactions and chemical reaction procedures to embeddings in a shared embedding space. Each tower can include a transformer network, a pooling layer, a normalization layer, and a neural network. The model is trained with labeled data pairs that include a chemical reaction and the text of a chemical reaction procedure for that chemical reaction. New queries can locate chemical reaction procedures for performing a given chemical reaction as well as procedures for similar chemical reactions. The architecture and training of the model make it possible to perform semantic matching based on chemical structures. The model is highly accurate providing an average recall at K=5 of 95.9%.

Machine learning system with two encoder towers for semantic matching

The present disclosure relates to methods and systems for converting Portable Document Format (PDF) documents to LaTeX files. The methods and systems use machine learning models to identify and extract PDF portions of a PDF document. The methods and systems create a LaTeX file for the PDF document using the PDF portions extracted by the machine learning models. The methods and systems provide an output with the LaTeX file for the PDF document. The LaTeX file is used to perform different actions on the PDF document.

Converting a portable document format to a latex format

The present disclosure relates to systems, methods, and computer-readable media for utilizing a concept graphing system to determine and provide relationships between concepts within document collections or corpora. For example, the concept graphing system can generate and utilize machine-learning models, such as a sparse graph recovery machine-learning model, to identify less-obvious correlations between concepts, including positive and negative concept connections, as well as provide these connections within a visual concept graph. Additionally, the concept graphing system can provide a visual concept graph that determines and displays concept correlations based on the input of a single concept, multiple concepts, or no concepts.

Determining concept relationships in document collections utilizing a sparse graph recovery machine-learning model

A relevance system ranks a set of medical studies based on a relevance of each medical study in the set of medical studies to a patient profile. The relevance system includes a relevance model. The relevance model determines a relevance of each medical study to the patient profile based on a semantic relationship score, a concept relationship score, and a term-occurrence score. The semantic relationship score is a measure of a similarity in semantic meaning of a medical study and a patient profile. The concept relationship score is a measure of the closeness of medical concepts in a medical study to medical concepts in a patient profile. The term-occurrence score is a measure of occurrences of terms in a medical study that also appear in a patient profile and the statistical significances of the terms.

Leveraging deep contextual representation, medical concept representation and term-occurrence statistics in precision medicine to rank clinical studies relevant to a patient

A system for implementing object state distributions obtains a first object state distribution associated with a first image set. The first object state distribution includes a first plurality of object state vectors generated based upon (i) a first set of object type classifications associated with the first image set and (ii) a first set of image information based on the first image set. The system obtains a second object state distribution associated with a second image set. The second object state distribution comprises a second plurality of object state vectors generated based upon (i) a second set of object type classifications associated with the second image set and (ii) a second set of image information based on the second image set. The system determines a distance measure between the first object state distribution and the second object state distribution and assigns a label based upon the distance measure.

Systems and methods for utilizing object state distributions for image comparison tasks

A computer implemented method determines differences between documents. The method includes parsing a first document and a second document into respective distinct instances of content. The distinct instances of content are classified into different categories. Category specific matching algorithms are applied to each of the respective instances of content to determine a similarity score for each of the respective instances of content. Semantic differences between the first document and the second document are analyzed as a function of the similarity scores. A characterization of the semantic differences is generated.

Semantic difference characterization for documents

Machine learning to predict a layout type that each of a plurality of portions of a document appears in. This is done even though the computer-readable representation of the document does not contain information at the granularity of the prediction to be made that identifies which layout type that each of the plurality of document portions belongs in. For each of a plurality of the portions, the machine-learning system predicts the layout type that the respective portion appears in, and indexes the document using the predictions so as to result in a computer-readable index. The index represents a predicted layout type associated with each of the plurality of portions of the document. Thus, the index can be used to search based on position of a searched term within the document.

Machine-learning of document portion layout

The present disclosure relates to generating a complex entity index based on a combination of atomic and deep learned attributes associated with instances of a complex entity. For example, systems described herein generate a multi-dimensional representation of entity instances based on evaluation of digital content associated with the respective entity instances. Systems described herein further generate an index representation in which similarity of entity instances are illustrated and presented via an interactive presentation that enables a user to traverse instances of an entity to observe similarities and differences between instances of an entity that have similar embeddings to one another within a multi-dimensional index space.

Generating and presenting multi-dimensional representations for complex entities

Performing collaborative search engine searching. The method includes receiving user input at a user interface for performing a plurality searches on a first search engine. The method further includes receiving user input at the user interface applying one or more augmentation AI models to searches in the plurality of searches. The method further includes creating a shareable, executable package executable by one or more search engines based on the plurality of searches and the applied AI models that when executed by the search engines causes the search engines to apply the AI models to searches performed at the search engines.

Search engine functionality using shared AI models

Growing community of inventors

Redmond, WA, United States of America

Robin Abraham