| CPC G06F 16/9535 (2019.01) [G06F 16/90324 (2019.01); G06F 16/907 (2019.01); H04L 67/535 (2022.05)] | 20 Claims |
|
1. A system, comprising:
a processor; and
a non-transitory memory storing instructions that, when executed, cause the processor to:
receive, during a first session at a website displayed at a computing device associated with a user, a first search query from the user that provides information inputted by the user in a first native format associated with the computing device;
responsive to receiving the first search query, automatically perform a first process during the first session comprising:
format the first search query from the first native format to a structured format, wherein the structured format defines a machine readable encoding of a first text object that includes the information and different from the first search query, thereby producing a first structured query;
apply the first structured query to a first pre-trained model in a plurality of models to receive as output from the first pre-trained model a first query correction comprising one or more recommended items to be displayed on the website to the in response to the first text object, wherein the one or more recommended items are semantically identified by the first pre-trained model based on the first text object;
obtain engagement data responsive to displaying the first query correction at the computing device, the engagement data comprising at least one click by the user on the one or more recommended items indicative of at least a user rejection of the first query correction, a second search query different from the first search query and the first query correction, and/or and a user acceptation of the first query correction;
generate multi-dimensional feature data for the plurality of models based on item catalog data, historical user data, and the engagement data;
train the plurality of models to decrease a number of rejected query corrections based on the multi-dimensional feature data, thereby forming a first trained model from the first pre-trained model;
update, by at least one rule based model of the plurality of models and in accordance with the training based on the multi-dimensional feature data, one of a query-correction database and a typo-candidate database based at least in part on the engagement data, wherein the one of the query-correction database and the typo-candidate database is updated to include at least one candidate term selected from a title of an item associated with the at least one click and different from the at least one or more recommended items, and wherein the at least one candidate term satisfies a plurality of rules implemented by the rule based model, the plurality of rules comprising a first rule defining a predetermined maximum edit distance of a term in the search query and a second rule defining a threshold character match value, thereby representing a portion of the information, and is ranked on at least one of a translation probability, an n-gram score, or a penalty score of at least one candidate item;
receive, during a second session at the website displayed at another computing device associated with another user, a third search query from the other user that provides the information inputted by the other user in a second native format associated with the other computing device;
responsive to receiving the third search query, automatically perform a second process during the second session comprising:
format the third search query from the second native format to the structured format that defines the machine readable encoding of the candidate term, thereby producing the second structured query;
apply the second structured query to the first trained model to receive as output from the first trained model a second query correction for the third search query comprising the item to be displayed on the website to the other user wherein the item is semantically identified by the first trained model based on the first text object.
|