Context-aware Adversarial Training for Name Regularity Bias in Named Entity Recognition

TL;DR

This paper investigates name regularity bias in NER models, introduces a diagnostic testbed, and proposes a novel adversarial training method to reduce bias and improve context utilization.

Contribution

The paper introduces NRB, a new benchmark for diagnosing name regularity bias, and proposes a model-agnostic adversarial training approach to mitigate this bias in NER models.

Findings

All tested models exhibit name regularity bias.

Adversarial training significantly reduces bias and improves context-based predictions.

Combining adversarial training with data augmentation and parameter freezing yields further improvements.

Abstract

In this work, we examine the ability of NER models to use contextual information when predicting the type of an ambiguous entity. We introduce NRB, a new testbed carefully designed to diagnose Name Regularity Bias of NER models. Our results indicate that all state-of-the-art models we tested show such a bias; BERT fine-tuned models significantly outperforming feature-based (LSTM-CRF) ones on NRB, despite having comparable (sometimes lower) performance on standard benchmarks. To mitigate this bias, we propose a novel model-agnostic training method that adds learnable adversarial noise to some entity mentions, thus enforcing models to focus more strongly on the contextual signal, leading to significant gains on NRB. Combining it with two other training strategies, data augmentation and parameter freezing, leads to further gains.