EviDR: Evidence-Emphasized Discrete Reasoning for Reasoning Machine Reading Comprehension

Reasoning machine reading comprehension (R-MRC) aims to answer complex questions that require discrete reasoning based on text. To support reasoning, evidence, typically the concise textual fragments describe question-related facts, including topic entities and attribute values, are crucial clues from question answer. However, previous end-to-end methods achieve state-of-the-art performance rarely solve problem by paying enough emphasis modeling of missing opportunity further improve model’s ability for R-MRC. alleviate above issue, in this paper, we propose an Evidence-emphasized Discrete approach (EviDR), which sentence clause level evidence is first detected distant supervision, then used drive a module implemented with relational heterogeneous graph convolutional network derive answers. Extensive experiments conducted DROP (discrete over paragraphs) dataset, results demonstrate effectiveness our proposed approach. In addition, qualitative analysis verifies capability evidence-emphasized R-MRC (Code released at https://github.com/JD-AI-Research-NLP/EviDR).