Incorporating window-based passage-level evidence in document retrieval

This study investigated whether information retrieval can be improved if documents are divided into smaller subdocuments or passages, and the retrieval score for these passages are incorporated in the final retrieval score for the whole document. The documents were segmented by sliding a window of a certain size across the document. Each time the window stopped, it displayed/extracted a certain number of contiguous words. A retrieval score was calculated for each of the passages extracted, and the highest score obtained by a passage of that size was taken as the document’s “window score” for that window size. A range of window sizes were tried. The experimental results indicated that using a fixed window size of 50 gave better results than other window sizes for the TREC test collection. This window size yielded a significant retrieval improvement of 24% compared to using the wholedocument retrieval score. However, combining this window score and the wholedocument retrieval score did not yield a retrieval improvement. Identifying the highest window score for each document (using window sizes varying from 50 to 400 words), and adopting it as the document retrieval score yielded a retrieval improvement of about 5% over taking the size-50 window score. Different window sizes were found to work best for different queries. If we could predict accurately the best window size to use for each query, a maximum retrieval improvement of 42% could be obtained. However, an effective way has not been found for predicting which window size would give the best results for each query.