The System RELIEFS: A New Approach for Information Filtering

In this year's filtering track, we implemented a system called RELIEFS that tries to learn about the prediction capability of words or conjunctions of words for the relevance of documents. The novelty of the system resides in two main points. First, the features used in the prediction involve both : the implication D->Q (from document to query), and the reverse implication Q->D. This is different from usual approaches where only the first of the implication is used. Therefore, the relevance estimation of a document combines the probability that a document containing a term is relevant, and the reverse probability the probability that a term appears in relevant documents. The second novelty is that, in addition to the use of words as prediction elements, we also consider word combinations (conjunctions). However, not all combinations are significant. Therefore, an incremental algorithm is developped to select only the meaningful conjunctions. To limit the number of conjunctions, we do not use a cut on conjunction length. Rather, we eliminate the conjunctions A&B that bear the same information as A or as B. Our first results prove the feasibility of the approach. Other experiments are ongoing in order to fully evaluate this approach.