An Empirical Analysis of Multiple-Turn Reasoning Strategies in Reading Comprehension Tasks

Reading comprehension (RC) is a challenging task that requires synthesis of across sentences and multiple turns of reasoning. Using a-of-the-art RC model, we empirically investigate the performance of-turn and multiple-turn reasoning on the SQuAD and MS MARCO datasets. The model is an end-to-end neural network with iterative attention, and uses learning to dynamically control the number of turns. We find that-turn reasoning outperforms single-turn reasoning for all question and types; further, we observe that enabling a flexible number of turns improves upon a fixed multiple-turn strategy. %across all question, and is particularly beneficial to questions with lengthy, descriptive. We achieve results competitive to the state-of-the-art on these two.