Microsoft Technology Licensing, LLC

Textio, Inc.

Kevin William Humphreys

Jesse Clay Satterfield

Jensen Michael Harris

Saliha Azzam

David Neal Weise

Kevin Roland Powell

Michael V Calcagno

Dawn Marie Wright

Martin David McClellan

Laurie Lee Dermer

Orion Buckminster Montoya

Christopher William Harland

Olivia Ann Gunton

Douglas W Potter

Jason S Hamilton

An electronic input document presented on a display of a client is examined to identify a text unit in the electronic input document and contextual information about the input document. A set of annotations for the text unit and the input document are determined responsive to the contextual information for the text unit. Responsive to the set of annotations, a set of candidate texts are identified from a corpus of documents that can replace the text unit. The candidate texts are evaluated in the set of candidate texts to identify a subset of the set of candidate texts as a set of replacement texts for the text unit. At least one replacement text from the set of replacement texts is presented on the display of the client.

Passively suggesting text in an electronic document

A user can request suggestions for text to insert at a position in an electronic input document. A set of annotations is obtained from the input document and contextual information associated with the document. A set of candidate texts from source documents are identified for the suggested text and evaluated. Each evaluation may indicate a likelihood that the corresponding candidate text is suitable to insert at the requested position of the input document. One or more of the candidate texts are presented to the client device as the suggested text to insert at the requested position of the electronic document.

Suggesting text in an electronic document

Facts are extracted from electronic documents by recognizing factual descriptions using a fact-word table to match to words of the electronic documents. The words of those factual descriptions may be tagged with the appropriate part of speech. More detailed analysis is then performed on those factual descriptions, rather than on the entire electronic document, and particularly to the text in the neighborhood of the fact-word matches. The analysis may involve identifying the linguistic constituents of each phrase and determining the role as either subject or object. Exclusion rules may be applied to eliminate those phrases unlikely to be part of facts, the exclusion rules being based in part on the linguistic constituents. Scoring rules may be applied to remaining phrases, and for those phrases having a score in excess of a threshold, the corresponding sentence part, whole sentence, paragraph, or other document portion may be presented as representing one or more facts.

Distinguishing facts from opinions using a multi-stage approach

The present invention provides an interface and associated object model that exposes a comprehensive set of natural language processing features to an application developer. In one embodiment, the features include lexicon management services and proofing services.

Natural language processing interface

A method of identifying definitions in documents includes receiving text units as an input. Which of the text units includes a cue phrase is then identified. For text units identified as including a cue phrase, localized parsing is performed around the cue phrase to determine whether the text unit including the cue phrase contains a definition.

Definition extraction

A method of, and system for, generating a sentence from a semantic representation maps the semantic representation to an unordered set of syntactic nodes. Simplified generation grammar rules and statistical goodness measure values from a corresponding analysis grammar are then used to create a tree structure to order the syntactic nodes. The sentence is then generated from the tree structure. The generation grammar is a simplified (context free) version of a corresponding full (context sensitive) analysis grammar. In the generation grammar, conditions on each rule are ignored except those directly related to the semantic representation. The statistical goodness measure values, which are calculated through an analysis training phase in which a corpus of example sentences is processed using the full analysis grammar, are used to guide the generation choice to prefer substructures most commonly found in a particular syntactic/semantic context during analysis.

Growing community of inventors

Redmond, WA, United States of America

Kevin William Humphreys