US 12,147,499 B2
Machine-learning tool for generating segmentation and topic metadata for documents
Rajiv Jain, Vienna, VA (US); Varun Manjunatha, Potomac, MD (US); Joseph Barrow, College Park, MD (US); Vlad Ion Morariu, Potomac, MD (US); Franck Dernoncourt, Sunnyvale, CA (US); Sasha Spala, Newton, MA (US); and Nicholas Miller, Newton, MA (US)
Assigned to Adobe Inc., San Jose, CA (US)
Filed by Adobe Inc., San Jose, CA (US)
Filed on Sep. 5, 2023, as Appl. No. 18/242,075.
Application 18/242,075 is a continuation of application No. 17/091,403, filed on Nov. 6, 2020, granted, now 11,783,008.
Prior Publication US 2023/0409672 A1, Dec. 21, 2023
This patent is subject to a terminal disclaimer.
Int. Cl. G06F 18/214 (2023.01); G06F 16/33 (2019.01); G06F 18/21 (2023.01); G06F 18/2415 (2023.01); G06F 40/117 (2020.01); G06F 40/30 (2020.01); G06V 30/413 (2022.01)
CPC G06F 18/2148 (2023.01) [G06F 18/217 (2023.01); G06F 18/2415 (2023.01); G06F 40/117 (2020.01); G06F 40/30 (2020.01); G06V 30/413 (2022.01); G06F 16/33 (2019.01); G06V 2201/10 (2022.01)] 20 Claims
OG exemplary drawing
 
1. A method comprising:
accessing an input document;
accessing a trained segmentation-and-labeling model configured by an accompanying training parameter set to predict segment bounds and classify resulting predicted segments into one or more topics;
generating a predicted segment for the input document by applying a segmentation network of the trained segmentation-and-labeling model to an encoded text portion from the input document;
generating a topic for the predicted segment by applying a pooling network of the trained segmentation-and-labeling model to the predicted segment;
assembling, using the predicted segment and the topic, an output document including the input document, segment metadata identifying the predicted segment, and topic metadata identifying the topic;
storing or displaying the output document; and
reconfiguring one of the segmentation network or the pooling network using an updated training parameter set, while leaving the other of the segmentation network or the pooling network configured by original training parameters from the accompanying training parameter set.