Cognitive Graph for Multi-Hop Reading Comprehension at Scale

TL;DR

This paper introduces a cognitive graph framework for multi-hop question answering that combines implicit extraction and explicit reasoning, achieving high accuracy and explainability on large-scale web documents.

Contribution

It presents a novel cognitive graph approach inspired by dual process theory, enabling scalable, explainable multi-hop reasoning with state-of-the-art performance.

Findings

Achieved a joint F1 score of 34.9 on HotpotQA fullwiki dataset.

Outperformed previous best by a significant margin (23.6 F1).

Demonstrated effective handling of web-scale documents for complex reasoning.

Abstract

We propose a new CogQA framework for multi-hop question answering in web-scale documents. Inspired by the dual process theory in cognitive science, the framework gradually builds a \textit{cognitive graph} in an iterative process by coordinating an implicit extraction module (System 1) and an explicit reasoning module (System 2). While giving accurate answers, our framework further provides explainable reasoning paths. Specifically, our implementation based on BERT and graph neural network efficiently handles millions of documents for multi-hop reasoning questions in the HotpotQA fullwiki dataset, achieving a winning joint $F_1$ score of 34.9 on the leaderboard, compared to 23.6 of the best competitor.