Back to results
US 10,592,607 B2Granted
Iterative alternating neural attention for machine reading
AssigneeMicrosoft Technology Licensing, LLC
InventorsAlessandro Sordoni, Philip Bachman, Adam Peter Trischler
Life of this patent
Estimating the term…
Abstract
Described herein are systems and methods for providing a natural language comprehension system (NLCS) that iteratively performs an alternating search to gather information that may be used to predict the answer to the question. The NLCS first attends to a query glimpse of the question, and then finds one or more corresponding matches by attending to a text glimpse of the text.
Claims20
A system, comprising:
- first processing circuitry;
- second processing circuitry;
- third processing circuitry operably connected to the first processing circuitry and the second processing circuitry;
- and
- one or more storage devices storing computer executable instructions that when executed by the first, the second, and the third processing circuitries, perform a method, the method comprising:
- processing, by the first processing circuitry, a question relating to a document to produce a query glimpse, the query glimpse comprising one or more entities from only a portion of the question;
- processing, by the second processing circuitry, one or more passages of text in the document to produce a text glimpse, the text glimpse comprising one or more entities that may be relevant to answering the question, wherein the one or more entities is from only a portion of the text;
- and
- processing, by the third processing circuitry, the query glimpse and the text glimpse to produce a result, the result comprising one of:
- a set of candidate answers, the set comprising one or more candidate answers;
- or
- information that is not relevant in predicting an answer to the question.
9 more elements
A method, comprising:
- processing a question relating to a document to produce a query glimpse, the query glimpse comprising one or more entities from only a portion of the question;
- processing one or more passages of text in the document to produce a text glimpse, the text glimpse comprising one or more entities that may be relevant to answering the question, wherein the one or more entities is from only a portion of the text;
- and
- processing the query glimpse and the text glimpse to produce a result, the result comprising one of:
- a candidate set of answers, the set comprising one or more candidate answers;
- or
- information that is not relevant in predicting an answer to the question.
4 more elements
A method, comprising:
- processing, at a first time-step, a question relating to a document to produce a query glimpse, the query glimpse comprising one or more entities in a portion of the question;
- processing, at the first time-step, one or more passages of text in the document to produce a text glimpse, the text glimpse comprising one or more entities in a portion of the text that may be relevant to answering the question;
- processing the query glimpse and the text glimpse to produce a result;
- determining if the result comprises a set of candidate answers;
- when the result comprises the set of candidate answers, storing the set of candidate answers;
- determining if all of the query and the text glimpses have been processed;
- based on a determination that all of the query and the text glimpses have not been processed, repeating, at each subsequent time-step, until all of the query and the text glimpses have been processed, the operations of:
- processing the question relating to the document to produce an additional query glimpse, the additional query glimpse comprising one or more different entities in a different portion of the question;
- processing one or more additional passages of text in the document to produce an additional text glimpse, the additional text glimpse comprising one or more different entities in a different portion of the text that are relevant to answering the question;
- processing the additional query glimpse and the additional text glimpse to produce an additional result;
- determining if the additional result comprises another set of candidate answers;
- and
- when the result comprises the additional set of candidate answers, storing the additional set of candidate answers;
- and
- based on a determination that all of the query and the text glimpses have been processed, determining a predicted answer based on each stored set of candidate answers.
12 more elements
Dependencies are inferred from the claim text.