Attentive pseudo-relevance feedback network for query categorization

Patent metadata

Field	Value
Publication number	US-11960555-B2
Application number	US-202217685130-A
Country	US
Kind code	B2
Filing date	Mar 2, 2022
Priority date	Mar 3, 2021
Publication date	Apr 16, 2024
Grant date	Apr 16, 2024

Abstract

A method of providing results from a search engine comprises generating an initial set based on a query received from a user; generating an attention value based on the query and on the initial set; applying the attention value to the initial set; and presenting a set of results in response to the query, wherein the results are generated according to the application of the attention value to the initial set.

First claim

What is claimed is:

1. A method of providing results from a search engine, the method comprising:
generating an initial set based on a query received from a user, the initial set corresponding to documents from the search engine;
determining, by a computing device, a vector representative of each document in the initial set of the query;
deriving, by the computing device, one or more fields corresponding to each document of the initial set;
determining, by a model of the computing device, field embeddings representative of each of the one or more fields corresponding to each document of the initial set comprising:
determining, by the model of the computing device, a word-level embeddings and character-level embeddings for each word in the field,
adjusting, by the model of the computing device, the word-level embeddings based on the character-level embeddings, and
determining, by the model of the computing device, the field embeddings based on the adjusted word-level embeddings,
wherein the adjusted word-level embeddings refine the field embeddings to compensate for morphological errors in the documents;

generating, by the model of the computing device, an attention value based on the vector representative of each document in the query and on the refined field embeddings of the initial set;
applying, by the computing device, the attention value to the initial set to generate a first set of results;
determining, by the computing device, a category based on an application of the attention value; and
presenting, by the computing device, a second set of results in response to the query;
wherein the second set of results comprises one or more documents of the initial set determined based on the attention values and on the refined field embeddings.

2. The method of claim 1 , wherein the attention value is a set attention value, and wherein generating the set attention value comprises:
generating a field attention value based on the field embeddings and the query;
generating a document attention value based on the field attention value and the query; and
generating the set attention value based on the document attention value.

3. The method of claim 1 , wherein applying the attention value to the initial set comprises:
determining document embeddings for each document in the initial set; and
applying the attention value to the document embeddings.

4. The method of claim 3 , wherein the second set of results comprise a subset of the initial set based on the category.

5. The method of claim 1 , wherein the set of results comprise at least one of:
a query recommendation;
a revised subset of documents from the initial set; or
a document of the revised subset of documents.

6. The method of claim 1 , wherein the initial set comprises one or more levels of abstraction, each document of the initial set being allocated to at least one level of abstraction of the one or more levels of abstraction,
wherein the vector representative of each document is based on data associated with each document at the at least one level of abstraction.

7. A system for providing results from a search engine, the system comprising:
a non-transitory, computer-readable medium storing instructions; and
a processor configured to execute the instructions to:
generate an initial set based on a query received from a user;
derive one or more fields corresponding to each document of the initial set;
determine, by a model, a field embeddings representative of each of the one or more fields comprising:
determining, by the model, a word-level embeddings and character-level embeddings for each word in the field,
adjusting, by the model, the word-level embeddings based on the character-level embeddings, and
determining, by the model, the field embeddings based on the adjusted word-level embeddings,
wherein the adjusted word-level embeddings refine the field embeddings to compensate for morphological errors in the documents;
generate, by the model, an attention value based on the query and on the field embeddings of the initial set;
apply the attention value to the initial set to generate a set of results;
determining a category based on an application of the attention value; and
present the set of results in response to the query;
wherein the set of results comprises one or more documents of the initial set determined based on the attention values and on the refined field embeddings.

8. The system of claim 7 , wherein the attention value is a set attention value, and wherein generating the set attention value comprises:
generating a field attention value based on the field embeddings and the query;
generating a document attention value based on the field attention value and the query; and
generating the set attention value based on the document attention value.

9. The system of claim 7 , wherein applying the attention value to the initial set comprises:
determining document embeddings for each document in the initial set; and
applying the attention value to the document embeddings.

10. The system of claim 9 , wherein the set of results comprise a subset of the initial set based on the category.

11. The system of claim 7 , wherein the set of results comprise at least one of:
a query recommendation;
a revised subset of documents from the initial set; or
a document of the revised subset of documents.

12. The system of claim 7 , wherein the initial set comprises one or more levels of abstraction, each document of the initial set being allocated to at least one level of abstraction of the one or more levels of abstraction,
wherein the one or more fields representative of each document is based on data associated with each document at the at least one level of abstraction.

13. A method of configuring a search engine, the method comprising:
generating an initial set of results from a query;
deriving a set of fields corresponding to each document of the initial set;
determining, by a neural network, field embeddings representative of the set of fields corresponding to each document of the initial set comprising:
determining, by the neural network, a word-level embeddings and character-level embeddings for each word in the set of fields,
adjusting, by the neural network, the word-level embeddings based on the character-level embeddings, and
determining, by the neural network, the field embeddings based on the adjusted word-level embeddings,
wherein the adjusted word-level embeddings refine the field embeddings to compensate for morphological errors in the documents;
extracting a subset of relevant documents from the initial set of results;
generate a set attention value based on the query and on the field embeddings of the subset of relevant documents;
generating a classification based on a combination of the set attention value and the subset of relevant documents; and
presenting a revised set of results based on the classification;
wherein the revised set of results comprises one or more documents of the initial set determined based on the set attention value and on the refined field embeddings.

14. The method of claim 13 , wherein generating the set attention value comprises:
generating a field attention value based on the query and a set of fields; and
generating a document attention value based on the query and on the field attention value,
wherein the set attention value is generated based on the document attention value.

15. The method of claim 14 , further comprising:
wherein the field attention value is based on the field embeddings.

16. The met

Assignees

• Home Depot Product Authority Llc

Inventors

• Ahmadvand Ali
• Zahiri Sayyed M
• Hughes Simon
• Al Jadda Khalifeh
• Kallumadi Surya
• Agichtein Eugene

Classifications

• G06F16/9535 Primary

  Search customisation based on user profiles and personalisation · CPC title

• G06F16/906

  Clustering; Classification · CPC title

• G06F16/9532

  Query formulation · CPC title

• G06F16/9538 Primary

  Presentation of query results · CPC title

• G06F16/3322

  using system suggestions (G06F16/3325 takes precedence) · CPC title