US 12,236,363 B2
Machine comprehension of unstructured text
Adam Trischler, Montreal (CA); Philip Bachman, Montreal (CA); Xingdi Yuan, Westmount (CA); Alessandro Sordoni, Montreal (CA); and Zheng Ye, North York (CA)
Assigned to Microsoft Technology Licensing, LLC, Redmond, WA (US)
Filed by Microsoft Technology Licensing, LLC, Redmond, WA (US)
Filed on Jun. 24, 2022, as Appl. No. 17/848,579.
Application 17/848,579 is a continuation of application No. 15/598,139, filed on May 17, 2017, granted, now 11,379,736.
Claims priority of provisional application 62/337,720, filed on May 17, 2016.
Prior Publication US 2022/0327407 A1, Oct. 13, 2022
This patent is subject to a terminal disclaimer.
Int. Cl. G06N 3/04 (2023.01); G06F 40/279 (2020.01); G06F 40/30 (2020.01); G06N 3/044 (2023.01); G06N 5/04 (2023.01)
CPC G06N 5/04 (2013.01) [G06F 40/279 (2020.01); G06F 40/30 (2020.01); G06N 3/044 (2023.01)] 20 Claims
OG exemplary drawing
 
1. A method, comprising:
receiving a Cloze-style question that relates to a text comprising one or more sentences;
processing, by a first neural network of first processing circuitry, the text and the Cloze-style question to produce a set of candidate answers to the Cloze-style question along with a probability of correctness for each candidate answer, wherein each candidate answer in the set of candidate answers is included in the text; and
processing, by a second neural network of second processing circuitry, the text and a set of hypotheses to determine a predicted answer, wherein the operation of processing comprises:
forming a hypothesis by inserting each candidate answer into the Cloze-style question, wherein each hypothesis is included in the set of hypotheses;
determining a semantic similarity between each sentence in the text and each hypothesis; and
outputting the candidate answer inserted in the hypothesis with the highest semantic similarity as a predicted answer for the Cloze-style question.