Multi-range Reasoning for Machine Comprehension

TL;DR

This paper introduces MRU, a fast and effective encoder for machine comprehension that captures long and short-term dependencies without recurrent layers, achieving competitive and state-of-the-art results on multiple datasets.

Contribution

The paper presents MRU, a novel multi-range gating encoder that improves sequence representation for machine comprehension without relying on recurrent or convolutional layers.

Findings

Outperforms DFN on RACE by 1.5%-6% without recurrence.

Achieves competitive results on SearchQA and NarrativeQA without LSTM/GRU.

Further improves performance when combined with BiLSTM architectures.

Abstract

We propose MRU (Multi-Range Reasoning Units), a new fast compositional encoder for machine comprehension (MC). Our proposed MRU encoders are characterized by multi-ranged gating, executing a series of parameterized contract-and-expand layers for learning gating vectors that benefit from long and short-term dependencies. The aims of our approach are as follows: (1) learning representations that are concurrently aware of long and short-term context, (2) modeling relationships between intra-document blocks and (3) fast and efficient sequence encoding. We show that our proposed encoder demonstrates promising results both as a standalone encoder and as well as a complementary building block. We conduct extensive experiments on three challenging MC datasets, namely RACE, SearchQA and NarrativeQA, achieving highly competitive performance on all. On the RACE benchmark, our model outperforms DFN (Dynamic Fusion Networks) by 1.5%-6% without using any recurrent or convolution layers. Similarly, we achieve competitive performance relative to AMANDA on the SearchQA benchmark and BiDAF on the NarrativeQA benchmark without using any LSTM/GRU layers. Finally, incorporating MRU encoders with standard BiLSTM architectures further improves performance, achieving state-of-the-art results.