Similarity Computation in Novelty Detection and Biomedical Text Categorization

The novelty track was first introduced in TREC 2002. Given a TREC topic, the goal of this task in 2004 is to locate relevant and new information from a set of documents. From the results in TREC 2002 and 2003, we realized the major challenging issue of recognizing relevant sentences is the lack of information used in similarity computation among sentences. In this year, we utilized the method based on variants of employing an information retrieval (IR) system to find relevant and novel sentences. This methodology is called IR with reference corpus, which can also be considered as an information expansion of sentences. A sentence is considered as a query of a reference corpus, and similarity between sentences is measured in terms of the weighting vectors of document lists ranked by IR systems. Basically, relevant sentences are extracted by comparing their results on a certain information retrieval system. Two sentences are regarded as similar if their corresponding returned document lists by the IR system are similar. In novelty parts, we used similar approach to extract novel sentences from the sentences of the relevant part. An effectively dynamic threshold setting approach that is based on what percentage of relevant sentences is within a relevant document is presented. In this paper, we paid attention to three points: first, how to utilize the results of an IR system to compare the similarity between sentences; second, how to filter out the redundant sentences; third, how to determine appropriate relevance and novelty threshold.