Incremental Learning for Profile Training in Adaptive Document Filtering

In this paper, we describe our ideas and related experiments in TREC-11 Adaptive Filtering Track. In the track we focused much on a robust way for effective profile training. We developed an incremental learning method which selects pseudo positive documents in less bias from a few initial positive training documents. We also did some experiments with newly emerged information retrieval model, language model-based retrieval mechanism, to evaluate its performance when used in adaptive filtering task. Related experiment results show the incremental learning method can be helpful for profile training, while the new language model perform not well.