Comparing Document Segmentation Strategies for Passage Retrieval in Question Answering

Information retrieval (IR) techniques are used in question answering (QA) to retrieve passages from large document collections which are relevant to answering given natural language questions. In this paper we investigate the impact of document segmentation approaches on the retrieval performance of the IR component in our Dutch QA system. In particular we compare segmentations into discourse-based passages and window-based passages with either fixed sizes or variable sizes. We also look at the effect of overlapping passages and sliding window approaches. Finally, we evaluate the different strategies by applying them to our question answering system in order to see the impact of passage retrieval on the overall QA accuracy.