Method and server for generating an enriched query
Methods and servers for generating an enriched query are disclosed. The server acquires an indication of a user query, employs a first model to generate a first candidate string and a second candidate string that are respective candidates for the enriched query; employs a second model to determine that the first candidate string and the second candidate string are a pair of similar candidate strings, and causes display of only one of the pair of similar candidate strings as the enriched query on the electronic device. The second model is trained in a pair-wise configuration based on a training set. The training set includes a first and second training string and a label. The label is indicative of a similarity between the first training string with the second training string. The label is determined based on a comparison of training document(s) relevant to the first and second training string.
1. A method of generating an enriched query for a user of an electronic device, the method being executed by at least one server communicatively coupled to the electronic device, the at least one server hosting a search engine, the method comprising:
acquiring, by the at least one server, an indication of a user query from the electronic device, the user query associated with query data;
generating, by the at least one server employing a first machine learning algorithm, a first candidate string and a second candidate string based on the query data, the first candidate string and the second candidate string being respective candidates for the enriched query;
determining, by the at least one server, and by inputting the first candidate string and the second candidate string into a second machine learning algorithm, that the first candidate string and the second candidate string are a pair of similar candidate strings, wherein the second machine learning algorithm was trained to predict a likelihood that a first string is similar to a second string, and wherein the second machine learning algorithm was trained in a pair-wise configuration based on a training set,
the training set including a first training string representative of a first training query, a second training string representative of a second training query, and a label,
the label being indicative of a similarity between the first training string with the second training string, the label having been determined based on a comparison of a first training document relevant to the first training query against a second training document relevant to the second training query; and
causing, by the at least one server, display of only one of the pair of similar candidate strings as the enriched query on the electronic device.
2. The method of claim 1 , wherein the label is indicative that the first training string is similar to the second training string if the first training document matches the second training document.
3. The method of claim 1 , wherein the first training document is a first plurality of training documents and the second training document is a second plurality of training documents,
the first plurality of training documents containing top N most relevant documents to the first training query,
the second plurality of training document containing top M most relevant to the second training query, and
the label being indicative that the first training string is similar to the second training string if the first plurality of training documents and the second plurality of training documents share at least a pre-determined number of training documents.
4. The method of claim 3 , wherein the top N most relevant documents to the first training query and the top M most relevant documents to the second training query are determined by the search engine in an off-line mode prior to training of the second machine learning algorithm.
5. The method of claim 1 , wherein the first model is a ranking model configured to generate a ranked list of candidate strings, the first candidate string and the second candidate string being ranked in the ranked list of candidate strings, the causing comprising:
causing, by the at least one server, display of a higher ranked candidate string amongst the first candidate string and the second candidate string as the enriched query on the electronic device.
6. The method of claim 1 , wherein the first machine learning algorithm is a Deep Semantic Similarity Model (DSSM).
7. The method of claim 1 , wherein the second machine learning algorithm is a DSSM.
8. The method of claim 1 , wherein the enriched query is an extended query that includes the query, the causing comprising:
causing, by the at least one server, display of the extended query on the electronic device without displaying a search engine results page (SERP) responsive to the query.
9. The method of claim 1 , wherein the enriched query is a reformulated query, the causing comprising:
causing, by the at least one server, display of a search engine results page (SERP) responsive to the query on the electronic device, the SERP including a portion with the reformulated query.
10. The method of claim 1 , wherein the enriched query is an extended query that includes the query, the causing comprising:
causing, by the at least one server, display of a search engine results page (SERP) responsive to the query on the electronic device, the SERP including a portion with the extended query.
11. A method of generating an enriched query for a user of an electronic device, the method being performed by at least one server communicatively coupled to the electronic device, the at least one server hosting a search engine, the method executable by the at least one server, the method comprising:
acquiring, by the at least one server, query data associated with a query submitted by the user;
generating, by the at least one server employing a model, a first candidate string and a second candidate string based on the query data and being respective candidates for the enriched query;
determining, by the at least one server, a first plurality of search results based on the first candidate string;
determining, by the at least one server, a second plurality of search results based on the second candidate string;
in response to the first plurality of search results and the second plurality of search results sharing at least a pre-determined number of search results,
determining, by the at least one server, that the first candidate string and the second candidate string are a pair of similar strings; and
causing, by the at least one server, display of only one of the pair of candidate strings as the enriched query on the electronic device.
12. The method of claim 11 , wherein the model is a ranking model configured to generate a ranked list of candidate strings, the first candidate string and the second candidate string being ranked in the ranked list of candidate strings, the causing comprising:
causing, by the at least one server, display of a higher ranked candidate string amongst the first candidate strings and the second candidate string as the enriched query on the electronic device.
13. The method of claim 11 , wherein the model is a DSSM.
14. The method of claim 11 , wherein the enriched query is an extended query including the query, the causing comprising:
causing, by the at least one server, display of the extended query on the electronic device without displaying a search engine results page (SERP) responsive to the query.
15. The method of claim 11 , wherein the enriched query is a reformulated query, the causing comprising:
causing, by the at least one server, display of a search engine results page (SERP) responsive to the query on the electronic device, the SERP including a portion with the reformulated query.
16. The method of claim 11 , wherein the enriched query is an extended query that includes the query, the causing comprising:
causing, by the at least one server, display of a search engine results page (SERP) responsive to the query on the electronic device, the SERP including a portion with the extended query.
17. A system for generating an enriched query for a user of an electronic device, the system communicatively coupled to the electronic device, the system hosting a search engine, the system comprising at least one processor and memory comprising executable instructions, which, when executed by the at least one processor, cause the system to:
acquire an indication of a user query from the electronic device, the user query associated with query data;
generate, by a first machine learning algorithm, a first candidate string and a second candidate string based on the query data, the first candidate string and the second candidate string being respective candidates for the enriched query;
determine, by inputting the first candidate string and the second candidate string into a second machine learning algorithm, that the first candidate string and the second candidate string are a pair of similar candidate strings, wherein the second machine learning algorithm was trained to predict a likelihood that a first string is similar to a second string, and wherein the second machine learning algorithm was trained in a pair-wise configuration based on a training set,
the training set including a first training string representative of a first training query, a second training string representative of a second training query, and a label,
the label being indicative of a similarity between the first training string with the second training string, the label having been determined based on a comparison of a first training document relevant to the first training query against a second training document relevant to the second training query; and
cause display of only one of the pair of similar candidate strings as the enriched query on the electronic device.
18. The system of claim 17 , wherein the label is indicative that the first training string is similar to the second training string if the first training document matches the second training document.
19. The system of claim 17 , wherein the first training document is a first plurality of training documents and the second training document is a second plurality of training documents,
the first plurality of training documents containing top N most relevant documents to the first training query,
the second plurality of training document containing top M most relevant to the second training query, and
the label being indicative that the first training string is similar to the second training string if the first plurality of training documents and the second plurality of training documents share at least a pre-determined number of training documents.
20. The system of claim 19 , wherein the top N most relevant documents to the first training query and the top M most relevant documents to the second training query are determined by the search engine in an off-line mode prior to training of the second machine learning algorithm.