Evaluating Metrics for Bias in Word Embeddings

Abstract

Over the last years, word and sentence embeddings have established as text preprocessing for all kinds of NLP tasks and improved the performances significantly. Unfortunately, it has also been shown that these embeddings inherit various kinds of biases from the training data and thereby pass on biases present in society to NLP solutions. Many papers attempted to quantify bias in word or sentence embeddings to evaluate debiasing methods or compare different embedding models, usually with cosine-based metrics. However, lately some works have raised doubts about these metrics showing that even though such metrics report low biases, other tests still show biases. In fact, there is a great variety of bias metrics or tests proposed in the literature without any consensus on the optimal solutions. Yet we lack works that evaluate bias metrics on a theoretical level or elaborate the advantages and disadvantages of different bias metrics. In this work, we will explore different cosine based bias metrics. We formalize a bias definition based on the ideas from previous works and derive conditions for bias metrics. Furthermore, we thoroughly investigate the existing cosine-based metrics and their limitations to show why these metrics can fail to report biases in some cases. Finally, we propose a new metric, SAME, to address the shortcomings of existing metrics and mathematically prove that SAME behaves appropriately.

Automatically Distinguishing People’s Explicit and Implicit Attitude Bias by Bridging Psychological Measurements with Sentiment Analysis on Large Corpora

“…Specifically, a study conducted by [25] quantifies the degree to which gender bias differs with the corpora used for the pre-trained model and fine-tuning with additional data. Schroder et al [26] proposes various metrics of embedding biases and compares their strengths and weaknesses. Garrido et al [27] explore biases in embeddings through the lens of geometric spaces, providing a fresh perspective on the subject.…”

Performance Implications of Using Unrepresentative Corpora in Arabic Natural Language Processing

“…In this paper, we aim to bridge this gap by studying the performance implications of using such unrepresentative, inorganic corpora (produced by template-based translation or automatic bots creation/generation) by intrinsically evaluating two main NLP upstream tasks: word representation and language modeling, using word analogy and fillmask evaluations, respectively, to capture syntactic and semantic relations between words. We purposely choose these intrinsic evaluations over extrinsic evaluations such as text classification or machine translation because many studies have shown that extrinsic and intrinsic evaluations' results are not consistently correlated, and the performance of NLP downstream tasks is always task-specific and can be significantly influenced by fine-tuning procedures (Faruqui et al, 2016;Schröder et al, 2021;Cao et al, 2022). We believe that evaluating NLP upstream tasks intrinsically will give us useful insights into the quality of the Arabic Wikipedia editions' corpora and show how the quality of corpora affects the performance of these NLP tasks.…”