The Birth of Bias: A case study on the evolution of gender bias in an English language model

By Oskar van der Wal, Jaap Jumelet, Katrin Schulz, Willem H. Zuidema. 2022

24 citations recorded by the source corpus.

Detecting and mitigating harmful biases in modern language models are widely recognized as crucial, open problems. In this paper, we take a step back and investigate how language models come to be biased in the first place. We use a relatively small language model, using the LSTM architecture trained on an English Wikipedia corpus. With full access to the data and to the model parameters as they change during every step while training, we can map in detail how the representation of gender develops, what patterns in the dataset drive this, and how the model's internal state relates to the bias in a downstream task (semantic textual similarity). We find that the representation of gender is dynamic and identify different phases during training. Furthermore, we show that gender information is represented increasingly locally in the input embeddings of the model and that, as a consequence, debiasing these can be effective in reducing the downstream bias. Monitoring the training dynamics, allows us to detect an asymmetry in how the female and male gender are represented in the input embeddings. This is important, as it may cause naive mitigation strategies to introduce new undesirable biases. We discuss the relevance of the findings for mitigation strategies more generally and the prospects of generalizing our methods to larger language models, the Transformer architecture, other languages and other undesirable biases.

Subfields: Natural Language Processing, Safety, Robustness & Interpretability

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Natural Language Processing

In Safety, Robustness & Interpretability

Citations in the corpus

Pythia: A Suite for Analyzing Large Language Models Across Training and Scaling (2023, methodology): “…Kirtane et al., 2022; Choenni et al., 2021), or the relationship between the corpus statistics and the measured bias (Bordia & Bowman, 2019; Van der Wal et al., 2022b), researchers have generally lacked the tools to study the role of the training data on the learning dynamics of bias in large…”
Undesirable biases in NLP: Averting a crisis of measurement (2022, background): “Moreover, it is also possible that some bias measures are better during different phases of the model training (Van der Wal et al., 2022).”
Handling Bias in Toxic Speech Detection: A Survey (2022, methodology): “Sampling Topic proximity [104], Data mixing (augmentation) [4] Annotation Disagreement deconvolution [39] Lexical Length sensitive sampling [24], Knowledge-based term generalization [5], Data iltering [120], Regularization [49], Multi-task learning [103], Ensemble modeling [120] Racial Annotation pr”