Dual Co-Matching Network for Multi-choice Reading Comprehension

TL;DR

The paper addresses multi-choice reading comprehension by introducing the Dual Co-Matching Network (DCMN), which jointly models all pairwise relationships among passage, question, and candidate answers in a bidirectional, symmetric fashion and fuses information with a gating mechanism. It encodes inputs with a BERT-based encoder and computes three pairwise matchings (P-Q, P-A, Q-A), aggregating them into a final triplet representation for candidate scoring via cross-entropy loss. Experiments on the RACE and ROCStories datasets show state-of-the-art performance, with the single model outperforming human turkers on RACE full data, and ablation studies confirming the importance of bidirectional matching and gating. The work demonstrates that considering complete triplet interactions and directional information significantly improves multi-choice MRC, suggesting a strong direction for future transformer-based triplet matching approaches.

Abstract

Multi-choice reading comprehension is a challenging task that requires complex reasoning procedure. Given passage and question, a correct answer need to be selected from a set of candidate answers. In this paper, we propose \textbf{D}ual \textbf{C}o-\textbf{M}atching \textbf{N}etwork (\textbf{DCMN}) which model the relationship among passage, question and answer bidirectionally. Different from existing approaches which only calculate question-aware or option-aware passage representation, we calculate passage-aware question representation and passage-aware answer representation at the same time. To demonstrate the effectiveness of our model, we evaluate our model on a large-scale multiple choice machine reading comprehension dataset (i.e. RACE). Experimental result show that our proposed model achieves new state-of-the-art results.

Figures (1)

• Figure 1: The framework of our model. P-Passage, Q-Question, A-Answer. $\oplus$ indicates the gated mechanism in Eq. \ref{['eq3']}.