An Empirical Study on the Fairness of Pre-trained Word Embeddings

Emeralda Sesari, Max Hort, Federica SarroNAACL 2022

TL;DR

This paper empirically investigates the fairness of 15 publicly available pre-trained word embedding models across three common training algorithms (GloVe, word2vec, fastText) using four established bias metrics. The study reveals that fastText models generally exhibit less bias, and smaller vector lengths tend to correlate with higher bias, offering practical guidance for selecting fairer word embeddings.

What Problem Does This Paper Address?

Word embeddings, powerful tools for Natural Language Processing (NLP), are often used without deep scrutiny of their inherent biases. These biases, inherited from the vast text corpora they are trained on, can perpetuate societal stereotypes and lead to unfair outcomes in downstream applications. While numerous studies have proposed methods to quantify bias in word embeddings, there's a need for a comprehensive empirical comparison of popular pre-trained models and their susceptibility to different types of bias. This paper aims to fill this gap by systematically evaluating the fairness of widely used pre-trained embeddings, helping practitioners make more informed choices.

Methods

The researchers conducted a systematic literature review to identify commonly used pre-trained word embedding models and bias metrics. They then designed an empirical study to evaluate 15 pre-trained word embeddings from GloVe, word2vec, and fastText. The fairness of these embeddings was assessed using four frequently employed bias metrics: WEAT (Word Embedding Association Test), SEMBIAS, DIRECT BIAS, and ECT (Embedding Coherence Test). The study specifically focused on intrinsic biases, meaning biases directly measurable within the embedding space itself, and investigated the impact of vector length on fairness by comparing embeddings trained on the same corpus but with different dimensionality.

Datasets

The study utilized publicly available pre-trained word embeddings. These embeddings were trained on various large-scale text corpora, including:

  • GloVe: Twitter (2 billion tweets), Wikipedia 2014 + Gigaword (6 billion tokens), Common Crawl (42 billion and 840 billion tokens).
  • word2vec: Google News (3 billion tokens).
  • fastText: Wikipedia 2017, UMBC Web Base and statmt.org News (16 billion tokens), Common Crawl (600 billion tokens).

The paper does not introduce new datasets but rather uses existing pre-trained models derived from these large corpora.

Key Insights and Ideas

  • fastText as the "Fairer" Champion: Across multiple bias metrics, fastText models consistently emerged as the least biased, suggesting its architecture or training methodology might be more robust against inheriting societal stereotypes.
  • Vector Length Matters: The study found a general trend where larger vector lengths tend to correlate with improved fairness, indicating that higher dimensionality embeddings are more expressive and less prone to biased associations.
  • Bias Metrics Aren't Uniform: The paper highlights that different bias metrics can yield varying results, emphasizing the importance of selecting metrics appropriate for the specific type of bias being investigated.
  • Pre-trained Embeddings are a Double-Edged Sword: While convenient, their widespread use necessitates a thorough understanding of their potential biases, as they can inadvertently propagate harmful stereotypes.

Key Results

  • WEAT: fastText models were found to be the least biased for 5 out of 7 WEAT tests. Specifically, fastText Wiki News and fastText Common Crawl (with subword information) showed the lowest bias scores in various racial, age, and gender bias tests.
  • SEMBIAS: All fastText embeddings outperformed others, achieving high semantic definition scores and low stereotypical/none scores. fastText trained on Common Crawl was identified as the least biased.
  • DIRECT BIAS: One GloVe model trained on Wiki Gigaword with a 300-dimensional vector was found to be the least biased. However, the general trend across GloVe models suggested that larger vector lengths led to lower bias.
  • ECT: Similar to SEMBIAS, fastText models, particularly fastText Common Crawl, showed the lowest bias scores.
  • Overall: fastText models performed best across three of the four most used bias metrics, with fastText Common Crawl being consistently identified as the least biased.

Limitations and Future Work

The authors acknowledge that the measures used only indicate the presence of specific biases and cannot definitively establish the absence of all biases. Future work could involve:

  • Deeper analysis of fairness tendencies and their correlation with explicit gender gaps and survey data.
  • Investigating the extent to which embeddings reproduce bias.
  • Replicating the study with manually trained models using different parameters on identical text corpora.
  • Exploring the fairness of contextual word embeddings (e.g., ELMO, BERT).
  • Investigating bias in word embeddings for languages with grammatical gender.

Summary updated

BibTeX
@inproceedings{sesari2022empirical,
  author = {Emeralda Sesari and Max Hort and Federica Sarro},
  title = {{An Empirical Study on the Fairness of Pre-trained Word Embeddings}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2022},
}
Figure