Discrete Reasoning Templates for Natural Language Understanding

TL;DR

This paper introduces a reasoning framework that decomposes complex questions into simpler subquestions using predefined templates, improving interpretability and performance on arithmetic reading comprehension tasks.

Contribution

It proposes a novel template-based reasoning approach that enhances interpretability and reduces supervision for complex question answering in reading comprehension.

Findings

Competitive performance on DROP dataset arithmetic questions

Interpretable reasoning process

Requires minimal supervision

Abstract

Reasoning about information from multiple parts of a passage to derive an answer is an open challenge for reading-comprehension models. In this paper, we present an approach that reasons about complex questions by decomposing them to simpler subquestions that can take advantage of single-span extraction reading-comprehension models, and derives the final answer according to instructions in a predefined reasoning template. We focus on subtraction-based arithmetic questions and evaluate our approach on a subset of the DROP dataset. We show that our approach is competitive with the state-of-the-art while being interpretable and requires little supervision