Workshop Program

	Thomas Hofmann Director of Engineering, Google Zurich European Engineering Centre, Switzerland
	Talk title: Beyond the Library - Organizing the World's Information Talk time: March 28 2007, 11:30

Talk abstract:
This talk will present challenges involved in building a Web search engine and will touch on questions of system and algorithm design, in particular as they involve large scale data processing and data mining.

	Nick Craswell Associate Researcher, Microsoft Research Cambridge, UK
	Talk title: Challenges in Web Information Retrieval Talk time: March 28 2007, 14:30

Talk abstract:
When building a Web search engine, we can benefit from core IR techniques, such as probabilistic ranking models and evaluation methods. But we also face problems that are not yet so well-studied in the field of IR. This talk explores several of these. For efficiency reasons, we need to crawl the web selectively. This raises an interesting query-independent ranking problem. We have large-scale logs of user behaviour. I will present a novel approach for dealing with sparsity of this data. We may also have relevance judgments for a large number of queries, as in the new TREC "million query" track, which allows for large-scale parameter tuning experiments. Each of these problems lends itself to data-driven solutions. The talk should thus give a flavour of the work that goes on in the area of commercial Web IR.

	Douglas W. Oard Associate Professor, University of Maryland, USA
	Talk title: But How Will They Use It? Thinking About Evaluation of Integrated Multilingual Multimedia Information Systems Talk time: March 28 2007, 16:30

Talk abstract:
Speech recognition and machine translation techniques are evolving rapidly, creating new opportunities to build systems to support information seeking in large collections of multilingual and multimedia content. Little is presently known, however, about how people would use such systems to accomplish real tasks. In such circumstances, designers naturally rely on their own judgment to decide how component capabilities should be optimized and how those components should be integrated. Once that's been done, the next step is to put the resulting system in the hands of users in order to learn what they do with it.
In this talk, I will describe what we have learned so far from such a process. I'll begin by describing Rosetta, an integrated system that supports search and display of live and archived news feeds in four languages by users who know only English. I'll then introduce several user study designs that we have tried, several focused on formative evaluation of the Rosetta system, and one that was designed to support a summative comparison of four systems. I'll illustrate each study design with some examples of what we have learned to date from approach. The talk will conclude with a few remarks on what we see as next steps in this continuing process. This is joint work with research teams at IBM Research, the University of Pittsburgh, Carnegie Mellon University and the University of Maryland.

	Josef Sivic Research Associate, University of Oxford, UK
	Talk title: Efficient visual search and automatic annotation of videos Talk time: March 29 2007, 9:00 The presentation can be freely downloaded [ size: 11 MB, requires Acrobat Reader]

Talk abstract:
Despite the recent success of text based search (e.g. Google), visual search in unannotated image and video collections remains a challenging problem. The imaged appearance of a particular object can change significantly due to changing camera viewpoint, illumination, or partial occlusion by other objects. Furthermore, there is additional variation, such as changing facial expression, when searching for particular people.
In the first part of the talk, I will describe our work on visual retrieval of particular objects and people in videos where the query is specified by an image of the object/person. A live visual search for objects and actors will be demonstrated on several feature length movies. In the second part of the talk, I will show how readily available text annotation such as movie shooting scripts and subtitles, together with visual facial appearance, can be used to automatically label occurrences of characters in TV or film footage.
Joint work with Andrew Zisserman and Mark Everingham.

	Djoerd Hiemstra Department of Computer Science, Database Group, University of Twente, The Netherlands
	Talk title: PF/Tijah - Text Search in an XML Database System Talk time: March 28 2007, 14:00

Talk abstract:
PF/Tijah (Pathfinder/Tijah, pronounce as "Pee Ef Teeja") is a flexible open source text search system that is integrated with an XML/XQuery database management system. PF/Tijah implements several built-in XQuery functions and is therefore fully compliant with XQuery 1.0. It has a number of unique selling points that distinguishes it from other information retrieval systems:

• PF/Tijah supports retrieving arbitrary parts of the textual data, not just documents.
• PF/Tijah supports complex scoring and ranking of the retrieved results by means of structured queries that address both the document content and its structure.
• PF/Tijah supports ad hoc result presentation by means of its query language. For instance, when searching for a special issue of a journal, it is easy to print any information from that retrieval result on the screen in a declarative way (i.e., by means of its query language, not by means of a general purpose programming language).
• PF/Tijah supports text search combined with traditional database querying, including for instance joins on values. For instance, one could search for employees from the financial department that also worked for the sales department and that sent an email about "tax refunds"