NeuroCOLT workshop
on
Applications of Learning to Text and Images
Windsor, 30 April - 2 May 2001
Cumberland Lodge

Probabilistic Sequence Models and Information Retrieval

Hugo Zaragoza

hugoz@microsoft.com

Download Talk

As textual data becomes readily available, there is an increasing need for specialised tools capable of retrieving and manipulating this data. Information retrieval remains the crucial technology behind most tools, but for many novel tasks retrieval has to integrate complex sources of information and structure, and be directed towards specific goals going beyond the mere retrieval or classification of documents. In recent years machine learning has approached information retrieval tasks from two perspectives: i) applying recent developments in symbolic and statistical learning to improve the performance of classifiers in standard information retrieval tasks, and ii) extending the domain of applicability of information retrieval. The first perspective has been the most popular and has yielded impressive results for those tasks closest to the statistical classification setting. The second perspective is gaining increasing attention as information and needs become increasingly complex; I will concentrate on this point. In the first part of my talk I will give an overview of the evolution of information retrieval tasks, needs and data, and review the most relevant machine learning approaches to information retrieval, pointing out the existence of a gap between these two. I will then centre the discussion on probabilistic sequence models and the different applications they have found in information retrieval in the last few years. The generative nature of these models allows us to describe a wide range of tasks, from information extraction and summarisation to document classification. It also poses severe limitations on the nature of models considered. I will discuss several of these approaches and their applications.