Object ranking

Abstract

Object ranking is an emerging discipline within information retrieval that is concerned with the ranking of objects, e.g. named entities and their attributes, in context of given a user query, or application. In this tutorial we will address the different aspects involved when building an object ranking system. We will present the state-of-the-art research in object ranking, as well as going into detail about our hands-on experiences when designing and developing the system for object ranking as it is in production at Yahoo! today. This allows for a unique mixture of research and development that will give the participants in-depth insights into the problem of object ranking. The focus of current Web search engines is to retrieve relevant documents on the Web, and more precisely documents that match with the query intent of the user. Some users are looking for specific information, while other just want to access rich media content (images, videos, etc.) or explore a topic. In the latter scenario, users do not have a fixed or pre-determined information need, but are using the search engine to discover information related to a particular object of interest. In this scenario one can say that the user is in a exploratory mode. To support users in their exploratory search the search engines are offering semantic search suggestions. In this tutorial, we will present a generic framework for ranking related objects. This framework ranks related entities according to two dimensions: a lateral dimension and a faceted dimension. In the lateral dimension, related entities are of the same nature as the entity queried (e.g. Barcelona and Madrid, or Angelina Jolie and Jessica Alba). In the faceted dimension, related entities are usually not of the same type as the queried entity, and refer to a specific aspect of the queried entity (e.g. Jennifer Aniston and the tvshow Friends). In this tutorial we will describe the process of building a Web-scale object ranking system. In particular we will address the construction of a knowledge base that forms the basis for the object ranking, and the generation of ranking features using external sources such as search engine query logs, photo annotations in Flickr, and tweets on Twitter. Next, we will discuss machine learned ranking models using an ensemble of pair-wise preference models, and address various aspects of object ranking, including multi-media extensions, vertical solutions, attribute-aware ranking, and the importance of freshness. Last but not least, we will address the evaluation methodologies involved to tune the performance of Web-scale object ranking strategies.