A Study on Efficiency, Accuracy and Document Structure for Answer Sentence Selection

An essential task of most Question Answering (QA) systems is to re-rank the\nset of answer candidates, i.e., Answer Sentence Selection (A2S). These\ncandidates are typically sentences either extracted from one or more documents\npreserving their natural order or retrieved by a search engine. Most\nstate-of-the-art approaches to the task use huge neural models, such as BERT,\nor complex attentive architectures. In this paper, we argue that by exploiting\nthe intrinsic structure of the original rank together with an effective\nword-relatedness encoder, we can achieve competitive results with respect to\nthe state of the art while retaining high efficiency. Our model takes 9.5\nseconds to train on the WikiQA dataset, i.e., very fast in comparison with the\n$\\sim 18$ minutes required by a standard BERT-base fine-tuning.\n