UMBC at TREC 12

We present the results of UMBC’s participation in the Web and Novelty tracks. We explored various heuristics-based link analysis approaches to the Topic Distillation task. For the novelty task we tried several methods for exploiting semantic information of sentences based on the SVD technique. We used SVD to expand the query and to filter redundant sentences. We also used a clustering algorithm that is also based on SVD. 1 Web Track Topic Distillation Task Web Retrieval has long been characterized as a web-traversal problem [6, 8]. We centered our efforts in the topic distillation task at this year’s TREC around this concept. As in last year’s approach, we used our CARROT II (C2) [4] agentbased Distributed Information Retrieval system to implement our approach and to observe if the task itself could be augmented by a distributed retrieval perspective. In a C2 system, an agent’s task is to maintain a collection of documents and handle all retrieval operations, including the operations of collection selection and results fusion, as applied to the domain of distributed retrieval.