Two-Step Feature Selection and Neural Network Classification for the TREC-8 Routing

At the Caisse des Dépôts et Consignations (CDC), the Agence France-Presse (AFP) news releases are filtered continuously according to the users' interests. Once a user has specified a topic of interest, a filter is customized to fit this user's profile. Until now, these filters would rely on rule-based methods, whose efficiency is proven [Vichot et al., 1999], but which require a large amount of work for each specific filter. This drawback can be avoided by using statistical methods which have the ability to learn from examples of relevant documents. Recently, we have developed a methodology for the AFP corpus. This paper presents its application to the TREC-8 corpus. For the TREC-8 routing, one specific filter is built for each topic. Each filter is a classifier trained to recognize the documents that are relevant to the topic. When presented with a document, each classifier estimates the probability for the document to be relevant to the topic for which it has been trained. Since the procedure for building a filter is topic-independent, the system is fully automatic. Therefore, we describe it for one topic; the procedure is repeated 50 times. By making use of a sample of documents that have previously been evaluated as relevant or not relevant to a particular topic, a term selection is performed, and a neural network is trained. Each document is represented by a vector of frequencies of a list of selected terms. This list depends on the topic to be filtered; it is constructed in two steps. The first step defines the characteristic words used in the relevant documents of the corpus; the second one chooses, among the previous list, the most discriminant ones. The length of the vector is optimized automatically for each topic. At the end of the term selection, a vector of typically 25 words is defined for the topic, so that each document which has to be processed is represented by a vector of term frequencies. This vector is subsequently input to a classifier that is trained from the same sample. After training, the classifier estimates for each document of a test set its probability of being relevant; for submission to TREC, the top 1000 documents are ranked in order of decreasing relevance.