A Self Consistency Based Reranking for Narrative Question Answering

The article presents a self-consistency based reranking framework for narrative question answering (NQA), addressing the limitations of single-output decoding in pretrained language models. The method generates multiple candidate answers for each story-question pair and selects the final response based on semantic agreement, enhancing robustness without altering the underlying architecture. Evaluated on the NarrativeQA dataset, the approach shows significant performance improvements, particularly with FLAN-T5-Base (from 82.32% to 86.66%) and Pegasus-Large (from 72.50% to 87.07%), indicating its potential for practitioners to enhance answer consistency in NQA tasks.