Weighted Global Normalization for Multiple Choice Reading Comprehension\n over Long Documents

Motivated by recent evidence pointing out the fragility of high-performing\nspan prediction models, we direct our attention to multiple choice reading\ncomprehension. In particular, this work introduces a novel method for improving\nanswer selection on long documents through weighted global normalization of\npredictions over portions of the documents. We show that applying our method to\na span prediction model adapted for answer selection helps model performance on\nlong summaries from NarrativeQA, a challenging reading comprehension dataset\nwith an answer selection task, and we strongly improve on the task baseline\nperformance by +36.2 Mean Reciprocal Rank.\n