Introduction
Machine Reading Comprehension (MRC) models have achieved remarkable success, often matching human performance on standard benchmarks. However, a critical question remains: are these models truly "reading" and reasoning, or are they simply exploiting statistical shortcuts and memorized entity associations? This paper investigates the robustness of state-of-the-art models to Entity Renaming, a common real-world scenario where facts evolve and new entities emerge.
A robust model should identify that "Michael" won a race based on the sentence "Michael beat Jack," regardless of whether it has seen the name "Michael" before. If a model relies too heavily on memorized knowledge (e.g., associating specific names with specific roles), it becomes brittle when faced with distribution shifts.
Methodology: The Perturbation Pipeline
To evaluate this robustness, the authors developed a four-step pipeline to systematically rename entities in test sets:
- Answer Entity Recognition: Using spaCy to identify entities within gold answers.
- Perturbable Span Identification: Assigning metadata like gender or specific types (e.g., "City" vs. "Country").
- Candidate Name Sampling: Selecting replacements from three sources:
InDistName(names from the same test set),DBName(external databases like U.S. Census data), andRandStr(random alphabetical strings). - Name Substitution: Consistently swapping the entity name throughout the passage, question, and answer key to maintain contextual integrity.
Key Insights
The study evaluated BERT, RoBERTa, and SpanBERT across five datasets, including human-annotated (SQuAD, NQ, HotpotQA) and distantly supervised (SearchQA, TriviaQA) sets.
- Distant Supervision and Brittleness: Models trained on SearchQA and TriviaQA showed significantly larger performance drops (~20%) compared to human-labeled data (<6%). This suggests that automated labeling encourages models to memorize entity-answer associations rather than learning to extract evidence.
- The GPE Challenge: Renaming Geopolitical Entities (GPE) caused more significant failures than renaming people or organizations, likely due to the smaller vocabulary of locations encountered during training.
- SpanBERT's Robustness: SpanBERT consistently outperformed BERT and RoBERTa on renamed entities. Its pretraining objective—masking contiguous spans—forces the model to focus on the relationship between a span and its surrounding context, making it less sensitive to internal character strings.
Experimental Results
The impact of entity renaming is most pronounced in distantly supervised datasets and when using random strings, which represent the most extreme domain shift.
On TriviaQA, BERT’s Exact Match (EM) score dropped from 73.4% to 25.4% when names were replaced with random strings. In contrast, SpanBERT maintained a 51.9 EM in the same scenario, demonstrating superior structural understanding. This suggests that while all models lose "world knowledge" during renaming, SpanBERT is better at utilizing local context to find the correct answer span.
Conclusion
This research highlights a significant gap in the reasoning capabilities of current MRC models. While models appear proficient on standard benchmarks, their reliance on memorized entity names limits their generalization to new or evolving information. Moving forward, pretraining objectives like span-based masking and entity-aware training offer promising directions for building more robust and truly "comprehending" systems.
Summary updated