This work focuses on algorithms which learn from examples to perform multiclass text and speech categorization tasks. Our approach is based on a new and improved family of boosting algorithms. We describe in detail an implementation, called BoosTexter, of the new boosting algorithms for text categorization tasks. We present results comparing the performance of BoosTexter and a number of other text-categorization algorithms on a variety of tasks. We conclude by describing the application of our system to automatic call-type identification from unconstrained spoken customer responses.

Classic IR (information retrieval) is inherently predicated on users searching for information, the socalled "information need". But the need behind a web search is often not informational -- it might be navigational (give me the url of the site I want to reach) or transactional (show me sites where I can perform a certain transaction, e.g. shop, download a file, or find a map). We explore this taxonomy of web searches and discuss how global search engines evolved to deal with web-specific needs.

In its current implementation, the World-Wide Web lacks much of the explicit structure and strong typing found in many closed hypertext systems. While this property has directly fueled the explosive acceptance of the Web, it further complicates the already difficult problem of identifying usable structures and aggregates in large hypertext collections. These reduced structures, or localities, form the basis to simplifying visualizations of and navigation through complex hypertext systems. Much of the previous research into identifying aggregates utilize graph theoretic algorithms based upon structural topology, i.e., the linkages between items. Other research has focused on content analysis to form document collections. This paper presents our exploration into techniques that harness both the topology and textual similarity between items as well as integrate new analyses based upon actual usage of the Xerox's WWW space. Linear equations and spreading activation models are employed to arrange Web pages based upon functional categories, node types, and relevancy. Keywords Information Visualization, World Wide Web, Hypertext.

The interaction of suffixing algorithms and ranking techniques in retrieval performance, particularly in an online environment, was investigated. Three general purpose suffixing algorithms were used for retrieval on the Cranfield 1400, Medlars, and CACM test collections, with no significant improvement in performance shown for any of the algorithms. A failure analysis suggested three modifications to ranking techniques: variable weighting of term variants, selective stemming depend-ing on query length, and selective stemming depending on term importance. None of these modifications im-proved performance. Recommendations are made re-garding the uses of suffixing in an online environment. introduction Traditional statistically based keyword retrieval systems have been the subject of experiments for over 30 years. The use of simple keyword matching as a basis for re-trieval can produce acceptable results, and the addition of ranking techniques based on the frequency of a given matching term within a document collection and/or within a given document adds considerable improvement (Sparck Jones, 1972; Salton, 1983). The conflation of word variants using suffixing al-gorithms was one of the earliest enhancements to statistical keyword retrieval systems (Salton, 1971), and has become so standard a part of most systems that many system descriptions neglect to mention the use of suffixing, or to identify the algorithm was used. Suffixing was originally done for two principle reasons: the large reduction in stor-age required by a retrieval dictionary (Bell, 1979), and the increase in performance due to the use of word variants. Recent research has been more concerned with perfor-mance improvement than with storage reduction.

this paper we argue that scientific articles require a different summarization strategy than, for instance, news articles. We propose a strategy which concentrates on the rhetorical status of statements in the article: Material for summaries is selected in such a way that summaries can highlight the new contribution of the source paper and situate it with respect to earlier work. We provide a gold standard for summaries of this kind consisting of a substantial corpus of conference articles in computational linguistics with human judgements of rhetorical status and relevance. We present several experiments measuring our judges' agreement on these annotations. We also present an algorithm which, on the basis of the annotated training material, selects content and classifies it into a fixed set of seven rhetorical categories. The output of this extraction and classification system can be viewed as a single-document summary in its own right; alternatively, it can be used to generate task-oriented and user-tailored summaries designed to give users an overview of a scientific field.

Existing audio tools handle the increasing amount of computer audio data inadequately. The typical tape-recorder paradigm for audio interfaces is inflexible and time consuming, especially for large data sets. On the other hand, completely automatic audio analysis and annotation is impossible using current techniques.

Conventional document retrieval systems (e.g., Alta Vista) return long lists of ranked documents in response to user queries. Recently, document clustering has been put forth as an alternative method of organizing retrieval results (Cutting et al. 1992). A person browsing the clusters can discover patterns that could be overlooked in the traditional presentation. This paper describes two novel clustering methods that intersect the documents in a cluster to determine the set of words (or phrases) shared by all the documents in the cluster. We report on experiments that evaluate these intersectionbased clustering methods on collections of snippets returned from Web search engines. First, we show that word-intersection clustering produces superior clusters and does so faster than standard techniques. Second, we show that our O(n log n) time phrase-intersection clustering method produces comparable clusters and does so more than two orders of magnitude faster than all methods tested. I...

Term cooccurrence data has been extensively used in document retrieval systems for the identification of indexing terms that are similar to those that have been specified in a user query: these similar terms can then be used to augment the original query statement. Despite the plausibility of this approach to query expan-sion, the retrieval effectiveness of the expanded que-ries is often no greater than, or even less than, the effectiveness of the unexpanded queries. This article demonstrates that the similar terms identified by cooc-currence data in a query expansion system tend to occur very frequently in the database that is being searched. Unfortunately, frequent terms tend to discrimi-nate poorly between relevant and nonrelevant docu-ments, and the general effect of query expansion is thus to add terms that do little or nothing to improve the dis-criminatory power of the original query.

The past decade has seen a tremendous growth in the amount of experimental and computational biomedical data, specifically in the areas of Genomics and Proteomics. This growth is accompanied by an accelerated increase in the number of biomedical publications discussing the findings. In the last few years there is a lot of interest within the scientific community in literature-mining tools to help sort through this abundance of literature, and find the nuggets of information most relevant and useful for specific analysis tasks. This paper

In current systems for browsing image collections, users are presented with sets of thumbnail images arranged in some default order on the screen. We are investigating whether it benefits users to have sets of thumbnails arranged according to their mutual similarity, so images that are alike are placed together. There are, of course, many possible definitions of "similarity": so far we have explored measurements based on low-level visual features, and on the textual captions assigned to the images. Here we describe two experiments, both involving designers as the participants, examining whether similarity-based arrangements of the candidate images are helpful for a picture selection task. Firstly, the two types of similarity-based arrangement were informally compared. Then, an arrangement based on visual similarity was more formally compared with a control of a random arrangement. We believe this work should be of interest to anyone designing a system that involves presenting sets of images to users. Keywords Image retrieval, information visualisation, evaluation.