Does Your Model Classify Entities Reasonably? Diagnosing and Mitigating Spurious Correlations in Entity Typing

Nan Xu, Fei Wang, Bangzheng Li, Mingtao Dong, Muhao Chen

Abstract

Modern entity typing models often achieve high performance on standard benchmarks by exploiting "spurious correlations"—linguistic shortcuts such as word overlaps or memorized entity names—rather than demonstrating a true understanding of semantic context. This paper systematizes these biases into six distinct categories and introduces a Counterfactual Data Augmentation (CDA) approach. By generating training samples that break these surface-level patterns, CDA forces models to prioritize contextual cues. Experimental results on the UFET and OntoNotes datasets demonstrate that CDA significantly improves model robustness and generalization, particularly on debiased test sets, without compromising performance on original benchmarks.

Introduction

Entity typing is the task of assigning semantic labels (e.g., person, hotel, or event) to specific mentions within a sentence. It serves as a critical component for information extraction and knowledge graph construction. While state-of-the-art (SOTA) models achieve impressive scores, there is growing evidence that they are often "right for the wrong reasons."

For instance, in the sentence "I stayed at Treasure Island," a model might predict the type island simply because the word "Island" appears in the name, ignoring the context "stayed at" which clearly indicates the entity is a hotel. This reliance on surface-level shortcuts makes models fragile when encountering unfamiliar entities or subtle phrasing. This paper aims to diagnose these failures systematically and provide a methodology for training more faithful models.

Diagnosing Spurious Correlations

The authors identify six specific categories of bias that plague entity typing models:

  1. Mention-Context Bias: An over-reliance on the mention's surface form, ignoring the surrounding context.
  2. Lexical Overlapping: Predicting a type because it shares a word with the mention (e.g., predicting day for the mention "Next day").
  3. Named Entity Bias: Memorizing specific famous entities from the training set rather than learning generalizable features.
  4. Pronoun Bias: Failing to resolve the semantic reference of a pronoun, leading to incorrect typing.
  5. Dependency Bias: Focusing on a modifier rather than the headword (e.g., in "car spoilers," typing the entity as vehicle instead of part).
  6. Overgeneralization: Defaulting to frequent, coarse labels (like person) due to class imbalance in the training data.

To detect these biases, the researchers utilize a "cloze test" approach with a Pre-trained Language Model (PLM) like RoBERTa. If the PLM can predict the correct type using only the mention without any context, that sample is flagged as potentially biased, indicating that the dataset provides a shortcut that bypasses the need for contextual understanding.

Counterfactual Data Augmentation (CDA)

To mitigate these biases, the authors propose Counterfactual Data Augmentation (CDA). The goal is to create "counterfactual" versions of training data that break the identified shortcuts, forcing the model to look at the full sentence for cues. The augmentation strategies include:

  • Lexical Debiasing: Replacing words in the mention with synonyms that do not overlap with the label.
  • Entity Debiasing: Swapping famous names for obscure or synthetic ones to prevent memorization.
  • Structural Debiasing: Truncating multi-word mentions to include only the headword, addressing dependency bias.

By training on a mixture of original and debiased samples, the model is incentivized to ignore surface-level coincidences and focus on the underlying semantic relationships.

Experimental Setup

The study utilizes two primary benchmarks:

  • UFET (Ultra-Fine Entity Typing): A challenging dataset featuring over 10,000 types, ranging from general to extremely specific.
  • OntoNotes: A standard dataset with 89 hierarchical types.

Crucially, the authors developed debiased versions of these test sets. These sets are specifically designed to evaluate how models perform when their typical shortcuts are removed or made misleading.

Results and Analysis

The experiments reveal significant insights into the current state of entity typing:

  • Fragility of SOTA: Even high-performing models like MLMET suffer massive performance drops—ranging from 40% to 90% in specific categories—when simple perturbations are applied to mentions in the test set.
  • Robustness Gains via CDA: Models trained with CDA showed substantial improvements on debiased test sets. For the MLMET model, CDA improved the relative F1 score by 11.81% on the debiased UFET set.
  • The Headword Problem: CDA proved exceptionally effective at teaching models basic syntax. On samples testing dependency bias, a BiLSTM model's performance improved by 160.5% after debiasing, indicating it had finally learned to identify the correct headword of a mention.
  • No Performance Tax: Unlike many debiasing techniques that degrade performance on "easy" or standard data, CDA maintained or slightly improved performance on the original UFET and OntoNotes test sets.

Conclusion

This research highlights a critical vulnerability in entity typing models: the tendency to rely on spurious correlations rather than context. By categorizing these biases and introducing Counterfactual Data Augmentation, the authors provide a path toward more robust and faithful NLP systems. While the study focuses on six bias types, future work is needed to expand this taxonomy and explore whether newer Entailment-based (NLI) or bi-encoder models share these same vulnerabilities.

Summary updated

Abstract

Entity typing aims at predicting one or more words that describe the type(s) of a specific mention in a sentence. Due to shortcuts from surface patterns to annotated entity labels and biased training, existing entity typing models are subject to the problem of spurious correlations. To comprehensively investigate the faithfulness and reliability of entity typing methods, we first systematically define distinct kinds of model biases that are reflected mainly from spurious correlations. Particularly, we identify six types of existing model biases, including mention-context bias, lexical overlapping bias, named entity bias, pronoun bias, dependency bias, and overgeneralization bias. To mitigate model biases, we then introduce a counterfactual data augmentation method. By augmenting the original training set with their debiased counterparts, models are forced to fully comprehend sentences and discover the fundamental cues for entity typing, rather than relying on spurious correlations for shortcuts. Experimental results on the UFET dataset show our counterfactual data augmentation approach helps improve generalization of different entity typing models with consistently better performance on both the original and debiased test sets.
BibTeX
@inproceedings{xu2022does,
  author = {Nan Xu and Fei Wang and Bangzheng Li and Mingtao Dong and Muhao Chen},
  title = {{Does Your Model Classify Entities Reasonably? Diagnosing and Mitigating Spurious Correlations in Entity Typing}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2022},
  doi = {10.18653/v1/2022.emnlp-main.592},
}
Figure