Coreference Reasoning in Machine Reading Comprehension

TL;DR

This paper highlights the challenges of coreference reasoning in machine reading comprehension, proposes a new dataset reflecting these challenges, and demonstrates improved model performance using natural coreference data.

Contribution

It introduces a methodology for creating more representative coreference reasoning datasets and shows how to enhance models with natural coreference data.

Findings

State-of-the-art models struggle with coreference reasoning.

New dataset better reflects coreference challenges in MRC.

Training with natural coreference data improves model performance.

Abstract

Coreference resolution is essential for natural language understanding and has been long studied in NLP. In recent years, as the format of Question Answering (QA) became a standard for machine reading comprehension (MRC), there have been data collection efforts, e.g., Dasigi et al. (2019), that attempt to evaluate the ability of MRC models to reason about coreference. However, as we show, coreference reasoning in MRC is a greater challenge than earlier thought; MRC datasets do not reflect the natural distribution and, consequently, the challenges of coreference reasoning. Specifically, success on these datasets does not reflect a model's proficiency in coreference reasoning. We propose a methodology for creating MRC datasets that better reflect the challenges of coreference reasoning and use it to create a sample evaluation set. The results on our dataset show that state-of-the-art models still struggle with these phenomena. Furthermore, we develop an effective way to use naturally occurring coreference phenomena from existing coreference resolution datasets when training MRC models. This allows us to show an improvement in the coreference reasoning abilities of state-of-the-art models. The code and the resulting dataset are available at https://github.com/UKPLab/coref-reasoning-in-qa.