IP Library Granted Patent US 12,651,279
Granted Patent B2
US 12,651,279 · App. 15/893,424 · Granted Jun 9, 2026

Method and system for matching query to ads using query subword vectors

Inventors: Erik Ordentlich (San Jose, CA); Andrew Feng (Cupertino, CA); Milind Rao (Palo Alto, CA); Jun Shi (Fremont, CA)
Assignee: YAHOO AD TECH LLC
G06Q30/0256G06F16/24542G06F16/9535G06N3/08G06Q30/0277
View Patent ↗
Loading inventors, assignments & file history…
Monitor This Case
Get email alerts when status or documents change.
Order Certified Copies
Most orders are placed with the USPTO same day — all within 24 business hours.
Order via The Patent Place →
Pre-filled with this patent's details
Quick Facts
Patent No.
US 12,651,279
App. No.
15/893,424
Granted
Jun 9, 2026
Kind
B2
Abstract

The present teaching relates to identifying content that matches a query. Train data include queries, advertisement, and hyperlinks associated with query sessions. A plurality of subwords for each of the queries in the training data are identified. A query/ads model is then trained by optimizing vectors associated the plurality of subwords for each of the queries, advertisements, and hyperlinks in the training data with respect to an objective function. At least one vector associated with each of the queries is derived based on the plurality of subword vectors in the query/ads model that represent the plurality of subwords of the query.

Claims (49)

1 . A method, implemented on a machine having at least one processor, storage, and a communication platform for training a query/ads model used for identifying content matching a query, comprising:

generating, by a model training engine, training data including positive and negative training samples derived based on previous query sessions, wherein each training sample in the training data comprises a query, one or more advertisements previously displayed with search results related to the query, and hyperlinks linking to the search results and one or more advertisements, wherein the positive training samples include at least one pair of words representing user events that appear within a predetermined number of temporally ordered user events in the query sessions and the negative training samples include pairs of words that are randomly selected from the query sessions;

reducing a number of parameters of the query/ads model by removing v vectors for the query, the one or more advertisements, and the hyperlinks to improve efficiency in training the query/ads model;

generating, for each of the query, the one or more advertisements, and the hyperlinks in the training data, a vector by:

identifying, by the model training engine, a plurality of subwords in the query, advertisement, or hyperlink, and

reducing a dimension for the vector by combining, via one of a convolutional neural network (CNN) and a recurrent neural network (RNN), a plurality of subword u vectors for the plurality of subwords to generate the vector for the query, the advertisement, or the hyperlink; and

training, by the model training engine via machine learning, the query/ads model, by optimizing the parameters of the query/ads model through maximizing a modeled probability with respect to the positive training samples and minimizing the modeled probability with respect to the negative training samples.

2 . The method of claim 1 , wherein the combining is further via one of:

one that performs component-wise averaging; or

one that performs component-wise average with average of pair wise component-wise product.

3 . The method of claim 1 , further comprising:

receiving the future query and a plurality of candidate advertisements identified based on the query;

identifying a set of subwords from the future query;

retrieving, from the query/ads model, vectors for the set of subwords and vectors for the plurality of candidate advertisements; and

selecting, from the plurality of candidate advertisements, at least one selected advertisement for the query based on the vectors for the set of subwords and the vectors for the plurality of candidate advertisements.

4 . The method of claim 3 , wherein the step of selecting is performed further based on a vector of the future query obtained based on the vectors for the set of subwords of the future query.

5 . A non-transitory, machine readable-medium having information recorded thereon for training a query/ads model used for identifying content matching a query, wherein the information, when read by the machine, causes the machine to perform the following:

generating training data including positive and negative training samples derived based on previous query sessions, wherein each training sample in the training data comprises: a query, one or more advertisements previously displayed with search results related to the query, and hyperlinks linking to the search results and one or more advertisements, wherein the positive training samples include at least one pair of words representing user events that appear within a predetermined number of temporally ordered user events in the query sessions and the negative training samples include pairs of words that are randomly selected from the query sessions;

reducing a number of parameters of the query/ads model by removing v vectors for the query, the one or more advertisements, and the hyperlinks to improve efficiency in training the query/ads model;

generating, for each of the query, the one or more advertisements, and the hyperlinks in the training data, a vector by:

identifying a plurality of subwords in the query, advertisement, or hyperlink, and

reducing a dimension for the vector by combining, via one of a convolutional neural network (CNN) and a recurrent neural network (RNN), a plurality of subword u vectors for the plurality of subwords to generate the vector for the query, the advertisement, or the hyperlink; and

training, via machine learning, the query/ads model, by optimizing the parameters of the query/ads model through maximizing a modeled probability with respect to the positive training samples and minimizing the modeled probability with respect to the negative training samples.

6 . The non-transitory, machine readable-medium of claim 5 , wherein the combining is further via one of:

one that performs component-wise averaging; or

one that performs component-wise average with average of pair wise component-wise product.

7 . The non-transitory, machine readable-medium of claim 6 , wherein the information, when read by the machine, further causes the machine to perform:

receiving the future query and a plurality of candidate advertisements identified based on the query;

identifying a set of subwords from the future query;

retrieving, from the query/ads model, vectors for the set of subwords and vectors for the plurality of candidate advertisements; and

selecting, from the plurality of candidate advertisements, at least one selected advertisement for the query based on the vectors for the set of subwords and the vectors for the plurality of candidate advertisements.

8 . The non-transitory, machine readable-medium of claim 7 , wherein the step of selecting is performed further based on a vector of the future query obtained based on the vectors for the set of subwords of the future query.

9 . A system having a processor, storage, and a communication platform capable of connecting to a network for training a query/ads model used for identifying content matching a query, comprising:

a subword vector identifier implemented by the processor and configured for

generating training data including positive and negative training samples derived based on previous query sessions, wherein each training sample in the training data comprises a query, one or more advertisements previously displayed with search results related to the query, and hyperlinks linking to the search results and one or more advertisements, wherein the positive training samples include at least one pair of words representing user events that appear within a predetermined number of temporally ordered user events in the query sessions and the negative training samples include pairs of words that are randomly selected from the query sessions,

reducing a number of parameters of the query/ads model by removing v vectors for the query, the one or more advertisements, and the hyperlinks to improve efficiency in training the query/ads model; and

generating, for each of the query, the one or more advertisements, and the hyperlinks in the training data, a vector by:

identifying a plurality of subwords in the query, advertisement, or hyperlink, and

reducing a dimension for the vector by combining, via one of a convolutional neural network (CNN) and a recurrent neural network (RNN), a plurality of subword u vectors for the plurality of subwords to generate the vector for the query, the advertisement, or the hyperlink; and

a query/ads model optimization engine implemented by the processor and configured for training, via machine learning, the query/ads model, by optimizing the parameters of the query/ads model through maximizing a modeled probability with respect to the positive training samples and minimizing the modeled probability with respect to the negative training samples.

10 . The system of claim 9 , wherein the combining is further via one of:

one that performs component-wise averaging; or

one that performs component-wise average with average of pair wise component-wise product.

11 . The system of claim 9 , further comprising the query/ads model based ad selector implemented by the processor and configured for:

receiving the future query and a plurality of candidate advertisements identified based on the query;

identifying a set of subwords from the future query;

retrieving, from the query/ads model, vectors for the set of subwords and vectors for the plurality of candidate advertisements; and

selecting, from the plurality of candidate advertisements, at least one selected advertisement for the query based on the vectors for the set of subwords and the vectors for the plurality of candidate advertisements, wherein

the step of selecting is performed further based on a vector of the query obtained based on the vectors for the set of subwords of the query.

Assignments (3)
CHANGE OF NAME Recorded Mar 22, 2022
From: VERIZON MEDIA INC.
To: YAHOO AD TECH LLC
Reel/Frame 059472/0001 →
ASSIGNMENT OF ASSIGNOR'S INTEREST Recorded Oct 26, 2020
From: OATH INC.
To: VERIZON MEDIA INC.
Reel/Frame 054258/0635 →
ASSIGNMENT OF ASSIGNOR'S INTEREST Recorded Feb 9, 2018
From: ORDENTLICH, ERIK; FENG, ANDREW; RAO, MILIND; SHI, JUN
To: OATH INC.
Reel/Frame 044886/0429 →
Continuity (1)
Related Publication 20190251595A1 · Aug 15, 2019
References Cited (28)
US 5960397A · Rahim · 1999 [cited by applicant]
US 9400955B2 · Garimella · 2016 [cited by examiner]
US 10540608B1 · Dirac et al. · 2020 [cited by applicant]
US 20070214114A1 · Liu et al. · 2007 [cited by applicant]
US 20110093331A1 · Metzler et al. · 2011 [cited by applicant]
US 20150278200A1 · He · 2015 [cited by examiner]
US 20160189204A1 · Ma et al. · 2016 [cited by applicant]
US 20170220575A1 · Hohwald et al. · 2017 [cited by applicant]
US 20170262449A1 · Venkataraman et al. · 2017 [cited by applicant]
US 20170286401A1 · He · 2017 [cited by examiner]
US 20180150466A1 · Paquet · 2018 [cited by examiner]
US 20180247226A1 · Forman et al. · 2018 [cited by applicant]
US 20180330273A1 · Hu et al. · 2018 [cited by applicant]
US 20190065460A1 · Xin et al. · 2019 [cited by applicant]
US 20190114348A1 · Gao et al. · 2019 [cited by applicant]
US 20200151395A1 · Cao · 2020 [cited by examiner]
Mihajlo Grbovic et al., Scalable Semantic Matching of Queries to Ads in Sponsored Search Advertising, SIGIR Jul. 2016, 375-384 (Year: 2016). [cited by examiner]
Office Action mailed May 21, 2020 in U.S. Appl. No. 15/893,321. [cited by applicant]
Office Action mailed Mar. 16, 2021 in U.S. Appl. No. 15/893,321. [cited by applicant]
Office Action mailed May 18, 2021 in U.S. Appl. No. 15/893,381. [cited by applicant]
Ordentlich et al “Network-Efficient Distributed Word2vec Training System for Large Vocabularies”, CIKM'16, Oct. 24-28, 2016 .(Year: 2016). [cited by applicant]
Zhai et al “DeepIntent: Learning Attentions for Online Advertising with Recurrent Neural Networks”, KDD'Aug. 13-17, 16, 2016.(Year: 2016). [cited by applicant]
Edizel et al“Deep Character-Level Click-Through Rate Prediction for Sponsored Search”, ACM, 2017. (Year: 2017). [cited by applicant]
Mihajlo Grbovic, Scalable Semantic Matching of Queries to Ads in Sponsored Search Advertising, 2016 (Yean 2016). [cited by applicant]
1 Office Action mailed Sep. 14, 2020 in U.S. Appl. No. 15/893,321. [cited by applicant]
Office Action mailed Mar. 24, 2023 in U.S. Appl. No. 15/893,381. [cited by applicant]
Notice of Allowance mailed Apr. 8, 2024 in U.S. Appl. No. 15/893,381. [cited by applicant]
Office Action mailed Oct. 18, 2023 in U.S. Appl. No. 15/893,381. [cited by applicant]