This paper describes a method for linear text seg- mc. ntation which is twice as accurate and over seven times as fast as the state-of-the-art (Reynar, 1998). Inter-sentence similarity is replaced by rank in the local context. Boundary locations are discovered by divisive clustering.

Event detection and tracking is a somewhat recent area of information retrieval research. The detection is about spotting new, previously unreported real-life events from online news-feed, while the tracking assigns documents to previously spotted events. We propose a new vector model consisting of four semantic classes from the documents: locations, proper names, temporal expressions and normal terms that are stored in designated subvectors. We also propose a new similarity measure based on utilizing semantic classes. Moreover, due to the vagueness of the concept of event, we run our experiments with several different definitions.

Topic Detection and Tracking is an event-based information organization task where online news streams are monitored in order to spot new unreported events and link documents with previously detected events. The detection has proven to perform rather poorly with traditional information retrieval approaches. We present an approach that formalizes temporal expressions and augments spatial terms with ontological information and uses this data in the detection. In addition, instead using a single term vector as a document representation, we split the terms into four semantic classes and process and weigh the classes separately. The approach is motivated by experiments.

On-line sources of news typically follow a particular pattern when presenting updates on a news event over time. First, they produce a preliminary report on the event, and later send out updates as the story evolves. There are two classes of readers accessing the latter stories - these who have read the original announcement and are familiar with the story background and those who are \joining" the thread at a later point in time. Because of the existence of the two clases of readers, news sources typically include in consequent stories some information that was already present in earlier stories. We discuss our approach to identifying such repeated pieces of information in news threads and show how this knowledge can help in generating userspeci c summaries of entire threads of articles. 1 Introduction To be able to generate summaries of threads of articles, it is important to do two things: identify which articles belong together (because they refer to the same event) and id...

This paper presents a new algorithm for domain independent linear text segmentation which is twice as accurate and over seven times as fast as the state-of-the-art [22]. The algorithm and statistical summarisation techniques were applied to a practical problem, improving document navigation for the visually disabled.

We investigate the extraction of linked-object representations (LORs) from text for use in topic tracking. LORs provide us a way to represent relationships between objects found in text. We show the use of naive coreference resolution during the extraction of objects does not provide improvement over the absence of coreference resolution. We investigate the creation of links, or relationships, between objects through closeness of the objects in text for small and large window sizes. We present a new algorithm, \centers", for document comparison and evaluate its eectiveness, showing that it approaches traditional cosine similarity when the window size is large. It gives dierent answers when the window size is small, but does not perform as well as cosine similarity where vectors contain words as components. Also, the \centers" algorithm is able to provide lower miss rates than when using vectors with LORs as components.