Method and system for rewriting a query

Patent metadata

Field	Value
Publication number	US-11086866-B2
Application number	US-201615130217-A
Country	US
Kind code	B2
Filing date	Apr 15, 2016
Priority date	Apr 15, 2016
Publication date	Aug 10, 2021
Grant date	Aug 10, 2021

Abstract

The present teaching relates to rewriting a query and providing search results. In one example, a plurality of queries is obtained. For each of the plurality of queries, one or more search results are identified. The one or more search results have been obtained in response to the query and have been previously selected by a user submitting the query. A plurality of titles is obtained. Each of the titles corresponds to one of the one or more search results with respect to one of the plurality of queries. A model is generated based on the plurality of queries and the plurality of titles. The model is to be used for rewriting a query.

First claim

We claim:

1. A method, implemented on a machine having at least one processor, storage, and a communication platform connected to a network for rewriting a query, the method comprising:
obtaining a plurality of queries;
identifying, for each of the plurality of queries, one or more search results that were previously selected by a user with respect to the query;
obtaining a plurality of titles each of which corresponds to each of the one or more search results;
generating a query rewrite model based on a plurality of query-title pairs, wherein the plurality of query-title pairs are generated based on the plurality of queries, the plurality of titles, and a degree of relevance between each of the plurality of queries and a corresponding one of the plurality of titles as well as word alignment candidates for each of the plurality of query-title pairs, wherein at least one null-word is inserted into at least one of the word alignment candidates of the plurality of query-title pairs, and wherein the degree of relevance is estimated based on interactions of the user with respect to the one or more search results related to the query;
receiving a new query;
automatically generating, based on the query rewrite model and the new query, one or more rewritten queries; and
obtaining a plurality of search results based on the one or more rewritten queries and the new query.

2. The method of claim 1 , wherein generating the one or more rewritten queries comprises:
determining segmentation candidates for the new query;
identifying one or more phrases for each segmentation candidate;
generating a translated phrase for each identified phrase based on the query rewrite model;
determining rewritten query candidates based on the translated phrases;
ranking the rewritten query candidates; and
generating the one or more rewritten queries from the rewritten query candidates based on the ranking.

3. The method of claim 2 , further comprising:
determining a first control level for similarity between queries;
determining a second control level for length of a rewritten query; and
determining a third control level for a quantity of stop-words in a query, wherein the rewritten query candidates are ranked based on the first, second and third control levels.

4. The method of claim 1 , further comprising:
generating parallel training data based on the plurality of query-title pairs, wherein the query rewrite model is generated based on the parallel training data via a machine learning algorithm.

5. The method of claim 4 , wherein generating the query rewrite model comprises:
determining the word alignment candidates for each of the plurality of query-title pairs;
extracting one or more phrase pairs from each of the plurality of query-title pairs based on each of the word alignment candidates;
determining a score for each of the one or more extracted phrase pairs; and
selecting at least one word alignment from the word alignment candidates based on the score determined for each of the one or more extracted phrase pairs corresponding to each of the word alignment candidates, wherein the query rewrite model is generated further based on extracted phrase pairs corresponding to the selected at least one word alignment.

6. The method of claim 1 , wherein:
the plurality of queries are obtained from one or more search logs; and
the one or more search results are identified based on a click graph generated based on the one or more search logs, wherein the click graph represents relationships between queries and documents in the one or more search logs, wherein the interactions with which the degree of relevance is estimated from are extracted from the click graph.

7. The method of claim 1 , wherein each of the plurality of titles represents a document or a Uniform Resource Locator (URL) corresponding to one of the one or more search results.

8. The method of claim 5 , wherein the one or more phrase pairs are further extracted from each of the plurality of query-title pair based on each word alignment candidate including the at least one null-word inserted therein.

9. A method, implemented on a machine having at least one processor, storage, and a communication platform connected to a network for providing search results, the method comprising:
receiving an original query for online content;
generating a first set of search results based on the original query;
generating a query rewrite model based on parallel training data comprising a plurality of query-title pairs, wherein each of the plurality of query-title pairs includes a query and a title corresponding to a search result that was previously selected by a user with respect to the query and one or more word alignment candidates for each of the plurality of query-title pairs, wherein each of the plurality of query-title pairs includes information regarding a degree of relevance between the query and the title, the degree of relevance being estimated based on interactions of the user with respect to the search result, wherein each of the one or more word alignment candidates aligns a word or phrase from a respective query to a word or phrase from respective title for each of the plurality of query-title pairs, and wherein at least one null-word is inserted into at least one of the word alignment candidates;
obtaining a rewritten query based on the original query and the query rewrite model;
generating a second set of search results based on the rewritten query;
generating a list of search results based on the first set of search results and the second set of search results; and
providing the list of search results in response to the original query.

10. The method of claim 9 , wherein the list of search results includes one or more search results from the second set of search results.

11. The method of claim 9 , wherein generating the list of search results comprises:
determining a first score for each search result in the first set of search results;
determining a second score for each search result in the second set of search results;
generating a combined score for each search result in the first set of search results and the second set of search results, wherein when the first set of search results and the second of search results both comprise a first search result, the combined score of the first search result is equal to a maximum of the first score and the second score of the first search result; and
ranking the first set of search results and the second set of search results based on their respective combined scores to generate a ranked list of search results, wherein the list of search results is generated based on the ranked list of search results.

12. A system having at least one processor, storage, and a communication platform connected to a network for rewriting a query, comprising:
a click graph generator/updater configured for:
obtaining a plurality of queries, and
identifying, for each of the plurality of queries, one or more search results that were previously selected by a user with respect to the query;

a query-title pair generator configured for obtaining a plurality of titles each of which corresponds to each of the one or more search results;
a query rewrite model learner configured for generating a query rewrite model based on a plurality of query-title pairs, wherein the plurality of query-title pairs are generated based on the plurality of queries, the plurality of titles, and a degree of relevance between each of the plurality of queries and a corresponding one of the plurality of titles as well as word alignment candidates for each of the plurality of query-title pairs, wherein at lea

Assignees

• Yahoo Holdings Inc
• Verizon Media Inc

Inventors

• Tang Jiliang
• Yin Dawei
• Deng Hongbo
• Daly Tim
• Tan Chao
• Langlois Jean-Marc
• Chang Yi

Classifications

• G06F16/24534 Primary

  Query rewriting; Transformation · CPC title

• G06N20/00

  Machine learning · CPC title

• G06F16/24578

  using ranking · CPC title