Weighting of Passages in Question Answering

Modern text retrieval systems employ text segmentation during the indexing of documents. We show that, rather than returning the passages to the user, significant improvements are achieved on the semantic text similarity task on question answering (QA) datasets by combining all passages from a document into a single result with an aggregate similarity score. Following an analysis of the SemEval-2016 and 2017 task 3 datasets, we develop a weighted averaging operator that achieves state-of-the-art results on subtask B and can be implemented into existing search engines. Segmentation in information retrieval matters. Our results show that paying attention to important passages by using a task-specific weighting method leads to the best results on these question answering domain retrieval tasks.