Abstract
Document-level relation extraction (DocRE) attracts more research interest recently. While models achieve consistent performance gains in DocRE, their underlying decision rules are still understudied: Do they make the right predictions according to rationales? In this paper, we take the first step toward answering this question and then introduce a new perspective on comprehensively evaluating a model. Specifically, we first conduct annotations to provide the rationales considered by humans in DocRE. Then, we conduct investigations and discover the fact that: In contrast to humans, the representative state-of-the-art (SOTA) models in DocRE exhibit different reasoning processes. Through our proposed RE-specific attacks, we next demonstrate that the significant discrepancy in decision rules between models and humans severely damages the robustness of models. After that, we introduce mean average precision (MAP) to evaluate the understanding and reasoning capabilities of models. According to the extensive experimental results, we finally appeal to future work to consider evaluating the understanding ability of models because the improved ability renders models more trustworthy and robust to be deployed in real-world scenarios. We make our annotations and code publicly available.
Introduction
Document-level relation extraction (DocRE) has seen significant performance gains, yet the underlying decision rules of these models remain opaque. A critical question persists: do these models truly understand the semantic logic of a document, or are they "Clever Hans" predictors exploiting statistical shortcuts? This paper investigates this discrepancy by comparing model rationales against human-annotated evidence.
As shown in Figure 1, while humans rely on semantic markers like "their nation" to link an athlete to a country, state-of-the-art (SOTA) models often focus on irrelevant tokens such as punctuation or random years.
Methodology and Datasets
To diagnose these reasoning processes, the authors introduced DocRED-HWE (Human-annotated Word-level Evidence), a dataset containing 7,342 relational facts with fine-grained, word-level rationales. Unlike previous sentence-level benchmarks, this allows for a precise evaluation of a model's "thought process."
The researchers employed three primary diagnostic tools:
- Feature Attribution (Integrated Gradients): Assigning importance scores to every input token to visualize model focus.
- RE-specific Attacks: Targeted perturbations, such as masking evidence words or substituting antonyms, to test semantic robustness.
- Mean Average Precision (MAP): A metric used to quantify the alignment between model-identified "important" words and human rationales.
Key Insights: Shortcuts and Biases
The investigation revealed that SOTA models frequently bypass complex reasoning in favor of "shortcuts":
- The "Narrow Scope" of Reasoning: Models can retain 85% of their performance using only entity names, largely ignoring the document's context.
- Position Discrimination: Models exhibit a bias toward tokens at specific document positions (e.g., the very beginning), regardless of their actual relevance.
- Collider Bias: Models often learn non-causal correlations between high-frequency function words and specific relations.
Figure 4 illustrates this phenomenon: tokens like periods, commas, and "[SEP]" markers are frequently assigned high importance across various relation types, suggesting the model is "gaming" the dataset's structure rather than reading the text.
Experimental Results
The results highlight a significant "understanding gap":
- Low Semantic Alignment: MAP scores for all tested SOTA models were below 10%, indicating that their internal rationales rarely match human logic.
- Fragility to Entity Changes: When entity names were replaced with out-of-distribution (OOD) names, performance plummeted. For instance, KD-RoBERTa’s F1 score dropped from 67.12% to 7.57%.
- The Antonym Paradox: In antonym substitution attacks, models failed to update their predictions over 90% of the time, proving a lack of true semantic comprehension.
Conclusion
While current DocRE models achieve high leaderboard scores, they lack the robust reasoning required for real-world deployment. By introducing DocRED-HWE and the MAP evaluation metric, this work provides a framework for building more trustworthy models that prioritize human-like understanding over statistical artifacts. Future research should focus on training methods that explicitly align model attention with human rationales.
Summary updated