US 11,704,551 B2
Iterative query-based analysis of text
Po-Sen Huang, Bellevue, WA (US); Jianfeng Gao, Woodinville, WA (US); Weizhu Chen, Kirkland, WA (US); and Yelong Shen, Bothell, WA (US)
Assigned to Microsoft Technology Licensing, LLC, Redmond, WA (US)
Filed by Microsoft Technology Licensing, LLC, Redmond, WA (US)
Filed on Jun. 30, 2017, as Appl. No. 15/639,705.
Claims priority of provisional application 62/407,153, filed on Oct. 12, 2016.
Prior Publication US 2018/0101767 A1, Apr. 12, 2018
Int. Cl. G06N 3/08 (2006.01); G06F 16/33 (2019.01); G06N 3/04 (2006.01); G06N 5/04 (2006.01); G06N 3/044 (2023.01)
CPC G06N 3/08 (2013.01) [G06F 16/3347 (2019.01); G06N 3/044 (2023.01); G06N 5/04 (2013.01)] 16 Claims
OG exemplary drawing
 
1. A system comprising:
at least one processor; and
one or more computer-readable storage media storing a neural network that is executable by the at least one processor to implement functionality comprising:
an internal state to represent a state of a query for information about text-based content, the internal state evolving as different portions of the text-based content are analyzed;
an attention vector to be applied to different portions of the text-based content to cause the internal state to iteratively evolve from an initial state denoting a first vector representation of the state of the query through a plurality of subsequent states as the different portions of the text-based content are analyzed; and
a termination gate to:
maintain the internal state, wherein a termination criterion comprises a change in the internal state;
evaluate the internal state and to terminate analysis of the text-based content in response to an occurrence of the termination criterion;
to cause the internal state to be output as an answer to the query in response to termination of the analysis;
apply a reward value to a termination state and generate a trained instance of the neural network; and
in response to the termination criterion having a false value, generate an updated attention vector, and feed the updated attention vector into a state network to update a next internal state.