Browse Prior Art Database

Method and System for Learning from Multi-Faceted Relevance for Vertical Search Ranking

IP.com Disclosure Number: IPCOM000215782D
Publication Date: 2012-Mar-12
Document File: 3 page(s) / 58K

Publishing Venue

The IP.com Prior Art Database

Related People

Belle Tseng: INVENTOR [+4]

Abstract

A method and system for learning from multi-faceted relevance for vertical search ranking is disclosed. The method and system proposes a multi-faceted relevance formulation in which holistic relevance is decomposed into multiple faceted relevance.

This text was extracted from a Microsoft Word document.
At least one non-text object (such as an image or picture) has been suppressed.
This is the abbreviated version, containing approximately 47% of the total text.

Method and System for Learning from Multi-Faceted Relevance for Vertical Search Ranking

Abstract

A method and system for learning from multi-faceted relevance for vertical search ranking is disclosed.  The method and system proposes a multi-faceted relevance formulation in which holistic relevance is decomposed into multiple faceted relevance.

Description

Disclosed is a method and system for learning from multi-faceted relevance for vertical search ranking.

A vertical search may include one or more of a local search and real time search.  For instance, a user may perform one of local search such as looking for an apartment near a location and real time search such as searching for breaking news regarding a topic.  For a vertical search, one or more facets may be considered.  Facets are different parameters under which a search results are listed and displayed to a user.  For example, if a user is looking for a restaurant in a specific location, then results displayed may correspond to one of a dining place that is exactly situated in the specific location and a dining place close to the specific location having good user reviews.  In the example, the dining place that is exactly situated in the specific location represents a matching facet, the dining place that is close to the specific location represents a distance facet and reviews of the dining place represent a reputation facet.

The one or more facets may get quantitatively aggregated with the help of supervised learning approaches to obtain a holistic relevance of a search result.  Since there are only few facets for a vertical search, minimal amount of training data with holistic training signals are needed to learn an aggregation function.  Moreover, holistic relative preferences signals can be transformed to holistic absolute labels.  This is important in ranking evaluation since the most popular metrics is based on the holistic absolute labels.  The method and system uses an editorial data set with relative preferences as the holistic relevance.  One or more approaches may be used to learn aggregation functions such as, but not limited to, a rule based method, a linear method, a gradient boosting decision tree and joint learning.

To learn an aggregation function, training data with holistic relevance may also be needed.  Training data may be collected when a search engine learns ranking.  The training data contains relevance labels for both the query and the results.  In accordance with the method and system, editors are requested to label a sample query and a result in compliance with an editorial guideline.  Thereafter, a holistic relevance question is asked to determine overall holistic relevance of the result by considering all related facets.  The holistic relevance is important for obtaining ranking of the search results.  The holistic relevance  may be one of absolute relevance labels and relative preferences.  To help editors assess the holistic relevance,...