Chains of Reasoning over Entities, Relations, and Text using Recurrent Neural Networks

TL;DR

This paper introduces a neural network approach that combines symbolic reasoning with deep learning to improve multi-hop inference over entities and relations, achieving significant accuracy gains on large-scale knowledge bases.

Contribution

It presents three key advances: joint reasoning about entities and relations, neural attention for multiple paths, and shared RNN parameters for logical composition.

Findings

25% error reduction on Freebase+ClueWeb

53% error reduction on sparse relations

84% error reduction on WordNet reasoning tasks

Abstract

Our goal is to combine the rich multistep inference of symbolic logical reasoning with the generalization capabilities of neural networks. We are particularly interested in complex reasoning about entities and relations in text and large-scale knowledge bases (KBs). Neelakantan et al. (2015) use RNNs to compose the distributed semantics of multi-hop paths in KBs; however for multiple reasons, the approach lacks accuracy and practicality. This paper proposes three significant modeling advances: (1) we learn to jointly reason about relations, entities, and entity-types; (2) we use neural attention modeling to incorporate multiple paths; (3) we learn to share strength in a single RNN that represents logical composition across all relations. On a largescale Freebase+ClueWeb prediction task, we achieve 25% error reduction, and a 53% error reduction on sparse relations due to shared strength. On chains of reasoning in WordNet we reduce error in mean quantile by 84% versus previous state-of-the-art. The code and data are available at https://rajarshd.github.io/ChainsofReasoning