US 12,307,382 B1
Neural taxonomy expander
Emaad Ahmed Manzoor, Pittsburgh, PA (US); Rui Li, San Jose, CA (US); Dhananjay Shrouty, San Francisco, CA (US); and Jurij Leskovec, Stanford, CA (US)
Assigned to Pinterest, Inc., San Francisco, CA (US)
Filed by Pinterest, Inc., San Francisco, CA (US)
Filed on Oct. 3, 2019, as Appl. No. 16/592,115.
Int. Cl. G06N 5/022 (2023.01); G06N 3/04 (2023.01); G06N 3/08 (2023.01)
CPC G06N 5/022 (2013.01) [G06N 3/04 (2013.01); G06N 3/08 (2013.01)] 20 Claims
OG exemplary drawing
 
1. A computer-implemented method for automatically locating a taxonomy candidate into an existing taxonomic hierarchy, the computer-implemented method comprising:
training a neural taxonomy expander according to the existing taxonomic hierarchy, wherein training the neural taxonomy expander includes:
generating a plurality of candidate/hypernym pairs, the plurality of candidate/hypernym pairs including positive candidate/hypernym pairs and negative candidate/hypernym pairs;
subdividing the plurality of candidate/hypernym pairs into a training set and a validation set;
repeatedly, until a predetermined loss threshold from processing the validation set is met:
training the neural taxonomy expander according to the training set until the predetermined loss threshold is met in processing the training set;
validating the neural taxonomy expander according to the validation set; and
updating the neural taxonomy expander according to results of the training upon a determination that the predetermined loss threshold for the validation set is not met; and
generating an executable neural taxonomy expander according to the trained neural taxonomy expander;
generating an embedding vector of the taxonomy candidate;
projecting the embedding vector of the taxonomy candidate into a taxonomic hyperspace according to a taxonomy projection by the executable neural taxonomy expander;
identifying a set of closest neighbors of the existing taxonomic hierarchy to the projected embedding vector of the taxonomy candidate;
determining a closest neighbor of the identified set of closest neighbors as an immediate parent of the taxonomy candidate; and
adding the taxonomy candidate into the existing taxonomic hierarchy as a child node of the determined immediate parent.