US 2009/0024607 A1 · QUERY SELECTION FOR EFFECTIVELY LEARNING RANKING FUNCTIONS

Abstract

A learning system for a search ranking function model may include a computer program that iteratively refines the model using new queries and associated documents from an unlabeled training set. The unlabeled training set may include a set of queries for which the associated documents have not been labeled as “relevant” or otherwise labeled. The new queries may be selected based on a similarity to and an accuracy of each neighbor from a labeled training set, such as a labeled validation set. Upon selection, the documents associated with the new queries may be labeled. The new queries and their associated documents may be accumulated into a labeled training set, such as a labeled training set, and a refined model may be learned based on the augmented labeled training set. The model may be iteratively refined until it is determined that the model is adequate.

Claims20

Dependent claims

A computer-implemented method for refining a ranking function model using an unlabeled training set of unlabeled queries and associated documents, a labeled training set of labeled queries and associated documents, and a labeled validation set of labeled queries and associated documents, the method comprising:

iteratively performing learning steps including:
selecting at least one new query from the unlabeled queries based on a similarity to and an accuracy of each labeled neighbor query in the labeled validation set;
labeling a set of documents associated with the newly selected query;
accumulating the newly labeled documents into the labeled training set;
learning a refined model based on the labeled training set;
and
determining whether the model is adequate;
and
storing the model.

6 more elements

Dependencies are inferred from the claim text.

Query selection for effectively learning ranking functions

Life of this patent

Query selection for effectively learning ranking functions

Life of this patent

Abstract

Claims20