Method and apparatus for determining search results, device and computer storage medium

Patent metadata

Field	Value
Publication number	US-10896188-B2
Application number	US-201916278651-A
Country	US
Kind code	B2
Filing date	Feb 18, 2019
Priority date	Jun 8, 2018
Publication date	Jan 19, 2021
Grant date	Jan 19, 2021

Abstract

The present disclosure provides a method and apparatus for determining search results, a device and a computer storage medium, wherein the method comprises: obtaining a current query, a historical query sequence of the current query and candidate search results of the current query; inputting the current query, the historical query sequence of the current query and candidate search results of the current query into a search result-sorting model, scoring the candidate search results according to the search result-sorting model, and determining search results corresponding to the current query; wherein the scores rendered by the search result-sorting model for the candidate search results are determined according to a similarity degree between an integrated vector representation of the current query and the historical query sequence of the current query and vector representations of the candidate search results. The present disclosure can provide the user with search results which more accurately reflect the user's demands.

First claim

What is claimed is:

1. A method of determining search results, wherein the method comprises:
obtaining a current query, a historical query sequence of the current query and candidate search results of the current query;
inputting the current query, the historical query sequence of the current query and candidate search results of the current query into a search result-sorting model, scoring the candidate search results according to the search result-sorting model, and determining search results corresponding to the current query; wherein the scores rendered by the search result-sorting model for the candidate search results are determined according to a similarity degree between an integrated vector representation of the current query and the historical query sequence of the current query and vector representations of the candidate search results.

2. The method according to claim 1 , wherein the historical query sequence of the current query comprises N queries before the current query in the same session, N being a positive integer.

3. The method according to claim 1 , wherein an integrated vector representation of the current query and the historical query sequence of the current query is determined in the following manner:
integrating the vector representation of the current query and the vector representation of the historical query sequence of the current query.

4. The method according to claim 3 , wherein the vector representation of the historical query sequence of the current query is determined in the following manner:
using an attention mechanism to perform weighted processing for vector representations of respective queries in the historical query sequence of the current query, to obtain a vector representation of the historical query sequence of the current query.

5. The method according to claim 1 , wherein the vector representation is determined in the following manner:
allowing input content to go through neural network processing and then obtaining a vector representation of the input content; wherein the input content includes the current query or queries in the historical query sequence of the current query.

6. The method according to claim 5 , wherein the neural network comprises: a Recurrent Neural Network RNN or a Convolutional Neural Network.

7. The method according to claim 1 , wherein the method further comprises:
displaying search results corresponding to the current query in a search result page of the current query.

8. The method according to claim 1 , wherein the search results include:
relevant entities or relevant webpages.

9. A method of training a search result sorting model, wherein the method comprises:
using a search log to obtain a training sample, the training sample comprising a sample query, a historical query sequence of the sample query, candidate search results of the sample query, and click situations of the candidate search results;
regarding the sample query, the historical query sequence of the sample query and the candidate search results of the sample query as input of a sorting model, regarding scores of the candidate search results rendered by the sorting model as output, and training the sorting model to achieve a preset training target; wherein the scores rendered by the sorting model for the candidate search results are determined according to a similarity degree between an integrated vector representation of the sample query and the historical query sequence of the sample query and vector representations of the candidate search results of the sample query; the training target includes: maximizing relevancy between click situations of the candidate search results and scores of candidate search results;
using the sorting model obtained from the training to obtain a search result-sorting model.

10. The method according to claim 9 , wherein the search results comprise: first-class search results and second-class search results;
the sorting model comprises a shared vector sub-model, a first sorting sub-model and a second sorting sub-model;
inputting the sample query and the historical query sequence of the sample query into the shared vector sub-model to obtain an integrated vector representation of the sample query and the historical query sequence of the sample query;
regarding the integrated vector representation and the candidate first-class search results of the sample query as input of the first sorting sub-model, the first sorting sub-model outputting scores of respective candidate first-class search results; determining the scores of the candidate first-class search results according to a similarity degree between the integrated vector representation of the sample query and the historical query sequence of the sample query and the vector representations of the candidate first-class search results of the sample query;
regarding the integrated vector representation and candidate second-class search results of the sample query as input of the second sorting sub-model, the second sorting sub-model outputting scores of respective candidate second-class search results; determining the scores of the candidate second-class search results according to a similarity degree between the integrated vector representation of the sample query and the historical query sequence of the sample query and the vector representations of the candidate second-class search results of the sample query;
training the first sorting sub-model and second sorting sub-model to achieve respective preset training targets; the training target of the first sorting sub-model includes: maximizing relevancy between click situations of the candidate first-class search results and scores of candidate first-class search results, and the training target of the second sorting sub-model includes: maximizing relevancy between click situations of the candidate second-class search results and scores of candidate second-class search results;
after completion of the training, using the shared vector sub-model and one of the first sorting sub-model and the second sorting sub-model, to obtain the search result sorting model.

11. The method according to claim 10 , wherein the training the first sorting sub-model and second sorting sub-model comprises:
during training iteration, randomly selecting one from the first sorting sub-model and second sorting sub-model for training each time; or alternatingly training the first sorting sub-model and second sorting sub-model during the training iteration.

12. The method according to claim 10 , wherein the first-class search results are relevant entities, and the second-class search results are relevant webpages.

13. The method according to claim 12 , wherein a training object of the first sorting sub-model comprises: minimizing a preset first loss function;
wherein the first loss function is determined by using a negative log likelihood function of each P(e + |c,q t );
where q t is the sample query, e + is a clicked candidate relevant entity of the sample query, and c is the historical query sequence of the sample query,

P
⁡

(

e
+

|
c

,

q
t

)

Assignees

• Beijing Baidu Netcom Sci & Tec

Inventors

• Huang Jizhou
• Zhang Wei
• Sun Yaming
• Wang Haifeng

Classifications

• G06N3/044

  Recurrent networks, e.g. Hopfield networks · CPC title

• G06N3/09

  Supervised learning · CPC title

• G06N3/0442

  characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU] · CPC title

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title

• G06F16/24578 Primary

  using ranking · CPC title