US 12,223,002 B2
Semantics-aware hybrid encoder for improved related conversations
Pinkesh Badjatiya, Madhya Pradesh (IN); Tanay Anand, Bangalore (IN); Simra Shahid, Uttar Pradesh (IN); Nikaash Puri, New Delhi (IN); Milan Aggarwal, Delhi (IN); S Sejal Naidu, Madhya Pradesh (IN); and Sharat Chandra Racha, Telangana (IN)
Assigned to ADOBE INC., San Jose, CA (US)
Filed by ADOBE INC., San Jose, CA (US)
Filed on Nov. 10, 2021, as Appl. No. 17/454,445.
Prior Publication US 2023/0143777 A1, May 11, 2023
Int. Cl. G06F 16/9536 (2019.01); G06F 16/9538 (2019.01); G06F 40/20 (2020.01)
CPC G06F 16/9536 (2019.01) [G06F 16/9538 (2019.01); G06F 40/20 (2020.01)] 16 Claims
OG exemplary drawing
 
1. A computer-implemented method of finding online relevant conversing posts, comprising:
receiving, by a web server serving an online forum, a query post on the online forum and a plurality of posts on the online forum;
generating a subject embedding and a body embedding for the query post, wherein the subject embedding is generated using a subject encoder based on a subject of the query post, and the body embedding is generated using a body encoder different from the subject encoder based on a body of the query post;
selecting, from the plurality of posts, a first number of posts based on the body embedding by computing, by a contextual similarity scoring module, a contextual similarity score between a conversing post of the online forum and the query post based on the body embedding for the query post;
selecting, from the first number of posts selected based on the body embedding, a second number of posts based on the subject embedding by computing, by a fine grained similarity scoring module, a fine grained similarity score between the query post and the conversing post based on the subject embedding for the query post, wherein the second number of posts is less than the first number of posts, wherein computing the fine grained similarity score between the subject embedding for the query post and the subject embedding for the conversing post comprises applying, by the subject encoder, a computer-implemented multi-lingual classifier to the subject of the query post and the subject of the conversing post and wherein the subject embedding for the conversing post is a first numeric vector, and the subject embedding for the query post is a second numeric vector, and calculating the fine grained similarity score comprises calculating an L2 distance between the first numeric vector and the second numeric vector; and
providing, by the web server, the conversing post in response to the query post based on the contextual similarity score and the fine grained similarity score.