[Paper Review] DuReaderrobust: A Chinese Dataset Towards Evaluating the Robustness of Machine Reading Comprehension Models.

Hongxuan Tang, Jing Liu+4 more|arXiv (Cornell University)|Apr 23, 2020

Topic Modeling23 references4 citations

TL;DR

This paper introduces DuReader_{robust}, a Chinese machine reading comprehension (MRC) dataset designed to evaluate model robustness across over-sensitivity, over-stability, and generalization challenges using natural text. Despite achieving human-level performance on standard benchmarks, state-of-the-art pre-trained MRC models significantly underperform on DuReader_{robust}, highlighting critical robustness gaps in real-world applications.

ABSTRACT

Machine Reading Comprehension (MRC) is a crucial and challenging task in natural language processing. Although several MRC models obtains human parity performance on several datasets, we find that these models are still far from robust. To comprehensively evaluate the robustness of MRC models, we create a Chinese dataset, namely DuReader_{robust}. It is designed to challenge MRC models from the following aspects: (1) over-sensitivity, (2) over-stability and (3) generalization. Most of previous work studies these problems by altering the inputs to unnatural texts. By contrast, the advantage of DuReader_{robust} is that its questions and documents are natural texts. It presents the robustness challenges when applying MRC models to real-world applications. The experimental results show that MRC models based on the pre-trained language models perform much worse than human does on the robustness test set, although they perform as well as human on in-domain test set. Additionally, we analyze the behavior of existing models on the robustness test set, which might give suggestions for future model development. The dataset and codes are available at \url{this https URL}

Motivation & Objective

• To address the lack of robustness evaluation in existing MRC models despite their strong performance on standard benchmarks.
• To create a realistic test set that challenges models on over-sensitivity, over-stability, and generalization using natural, real-world text.
• To provide a benchmark that reflects real-world MRC application challenges beyond standard in-domain performance.
• To analyze model behaviors on robustness challenges and guide future model development.
• To release a publicly available dataset and codebase for reproducible robustness evaluation in Chinese MRC.

Proposed method

• Construction of a new Chinese MRC dataset, DuReader_{robust}, by modifying existing DuReader samples to introduce robustness challenges while preserving natural language.
• Design of three robustness evaluation dimensions: over-sensitivity (to minor input perturbations), over-stability (to irrelevant or redundant text), and generalization (to out-of-domain or diverse contexts).
• Use of natural, human-written questions and documents to ensure real-world applicability, avoiding artificial or unnatural perturbations common in prior work.
• Evaluation of pre-trained MRC models (e.g., BERT-based) on both in-domain and DuReader_{robust} test sets to compare performance drops.
• Analysis of model failure patterns on the robustness test set to identify common failure modes and inform future model design.
• Public release of the dataset and code at a dedicated URL to support reproducibility and further research.

Experimental results

Research questions

• RQ1 How do state-of-the-art pre-trained MRC models perform on a robustness-focused test set using natural text?
• RQ2 To what extent do MRC models exhibit over-sensitivity to minor input variations in natural Chinese text?
• RQ3 How do models handle over-stability when faced with redundant or irrelevant information in documents?
• RQ4 Can existing MRC models generalize effectively to out-of-domain or diverse contexts in real-world scenarios?
• RQ5 What are the dominant failure patterns of current MRC models when evaluated on natural, robustness-challenging examples?

Key findings

• Pre-trained MRC models achieve near-human performance on standard in-domain test sets but show significant performance drops on the DuReader_{robust} benchmark.
• Models exhibit high over-sensitivity, failing on minor, natural variations in questions or documents that humans handle with ease.
• Over-stability is observed when models ignore relevant information or are misled by redundant text, indicating poor attention control.
• Generalization performance is notably poor, with models struggling on out-of-domain or diverse contextual inputs.
• The performance gap between humans and models on DuReader_{robust} is substantial, indicating a critical robustness deficit in current MRC systems.
• Failure analysis reveals consistent model behaviors, such as reliance on surface patterns and limited reasoning under perturbations, suggesting directions for future improvement.