Gendered Ambiguous Pronouns Shared Task: Boosting Model Confidence by Evidence Pooling

Sandeep AttreeACL 2019

TL;DR

This paper tackles the challenge of gender bias in coreference resolution systems, where models tend to be more confident in resolving masculine pronouns. They introduce a novel deep learning architecture, GREP, that effectively pools evidence from multiple off-the-shelf coreference models to boost confidence and accuracy, achieving state-of-the-art results on the Gendered Ambiguous Pronouns (GAP) task.

What Problem Does This Paper Address?

Coreference resolution, the task of identifying mentions in text that refer to the same real-world entity, is a fundamental NLP problem. However, existing state-of-the-art models exhibit a systematic gender bias, showing higher confidence when resolving masculine pronouns compared to feminine ones. This bias can lead to unfair or inaccurate outcomes in downstream applications. The Gendered Ambiguous Pronouns (GAP) shared task was created to address this issue by providing a dataset and benchmark for developing more robust and less biased coreference systems. The paper specifically aims to improve model confidence in resolving gendered ambiguous pronouns, particularly in challenging "Neither" cases where the pronoun might not refer to either of the provided candidate entities.

Methods

The authors propose a novel architecture called GREP (Gendered Resolution by Evidence Pooling). GREP builds upon a strong baseline model, ProBERT, which is essentially a fine-tuned BERT language model. ProBERT uses token embeddings from BERT's last layer, corresponding to the pronoun, to make a classification decision.

The key innovation is the Evidence Pooling module. This module leverages predictions from multiple existing, off-the-shelf coreference models. It treats these models as "black boxes" and uses their outputs as auxiliary evidence. The module employs a hierarchical attention mechanism:

  1. Mention-level aggregation: For each coreference model's predicted cluster, attention pooling is used to aggregate token embeddings of mentions within that cluster.
  2. Compatibility scoring: A transformer layer (using multi-head attention) is used to compute the compatibility between the pronoun's representation and the aggregated representations of the candidate entities (A and B) and the coreference model's cluster.
  3. Evidence vector creation: The compatibility scores are further aggregated across all coreference models to form a single "evidence vector."
  4. Final prediction: This evidence vector is concatenated with the pronoun's representation from ProBERT and fed into a final classification layer to produce the output probabilities.

This modular design allows for easy integration of new coreference solvers and other knowledge sources.

Datasets

The primary dataset used is the Gendered Ambiguous Pronouns (GAP) dataset, specifically the gap-development, gap-validation, and gap-test splits. The paper also mentions using a gpr-neither subset and a stage 2 test set from a Kaggle competition. The authors performed data sanitization on the dataset to correct ambiguous labels.

Key Insights and Ideas

  • Evidence Pooling is Key: The core insight is that by intelligently pooling evidence from multiple, potentially noisy, coreference solvers, GREP can overcome the limitations of individual models and significantly boost confidence and accuracy, especially for difficult "Neither" cases.
  • Leveraging Off-the-Shelf Models: The approach demonstrates the power of using existing, specialized NLP models as building blocks, even if their internal workings are not fully understood or directly trainable. This "black-box" integration is a practical way to enhance performance.
  • ProBERT's Surprising Effectiveness: The baseline ProBERT model, using only pronoun representation and mention tags, performs surprisingly well. The authors speculate this is due to BERT's attention mechanism specializing pronoun representations and dropout making it robust to noise.
  • Confidence Boost: GREP not only improves accuracy but also demonstrably increases model confidence, as shown by the probability distributions shifting closer to 1.0 for correct predictions.

Key Results

  • State-of-the-Art Performance: The GREP model achieved state-of-the-art performance on the GAP test data, reaching 92.5% F1 score with a gender bias ratio of 0.97.
  • Kaggle Competition Win: The end-to-end GREP solution placed 1st in the Kaggle competition for the Gendered Ambiguous Pronouns shared task, winning by a significant margin.
  • Significant Gains: Compared to the ProBERT baseline, GREP achieved improvements of 6 points in log loss and 2.8 points in F1 accuracy on the GAP test set.
  • Reduced Gender Bias: The ensemble models achieved performance very close to human performance with almost no gender bias.

Limitations and Future Work

  • Noisy Coreference Predictions: The Evidence Pooling module's performance is dependent on the quality of the off-the-shelf coreference models. The paper acknowledges that these models can be noisy, and their internal gradients are not propagated, limiting direct end-to-end optimization.
  • Computational Cost: Training and running multiple coreference models can be computationally expensive.
  • Future Work: The authors suggest extending GREP to solve gendered pronoun resolution beyond the gold-two-mention task, meaning without access to labeled gold spans. They also propose further analysis of ProBERT's attention heads to understand its surprising effectiveness.

Summary updated

BibTeX
@inproceedings{attree2019gendered,
  author = {Sandeep Attree},
  title = {{Gendered Ambiguous Pronouns Shared Task: Boosting Model Confidence by Evidence Pooling}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2019},
}
Figure