Debiasing Sentence Embedders through Contrastive Word Pairs

TL;DR

This work explores an approach to remove linear and nonlinear bias information for NLP solutions, without impacting downstream performance and compares this approach to common debiasing methods on classical bias metrics and on bias metrics which take nonlinear information into account.

Abstract

Over the last years, various sentence embedders have been an integral part in the success of current machine learning approaches to Natural Language Processing (NLP). Unfortunately, multiple sources have shown that the bias, inherent in the datasets upon which these embedding methods are trained, is learned by them. A variety of different approaches to remove biases in embeddings exists in the literature. Most of these approaches are applicable to word embeddings and in fewer cases to sentence embeddings. It is problematic that most debiasing approaches are directly transferred from word embeddings, therefore these approaches fail to take into account the nonlinear nature of sentence embedders and the embeddings they produce. It has been shown in literature that bias information is still present if sentence embeddings are debiased using such methods. In this contribution, we explore an approach to remove linear and nonlinear bias information for NLP solutions, without impacting downstream performance. We compare our approach to common debiasing methods on classical bias metrics and on bias metrics which take nonlinear information into account.

Figures (3)

• Figure 1: Number of pre-training steps per debiasing step vs detected bias. Higher Occupation task loss denotes less bias present in the embedding.
• Figure 2: Number of fine-tune steps per debiasing step vs detected bias. Higher bias classifier loss denotes less bias present in the embedding. The red line denotes the accuracy on the downstream task. The blue line shows the bias loss of the linear model and the yellow line denotes the bias loss of the nonlinear model. The bias scores and accuracies are averaged over 5 runs.
• Figure 3: DeepView visualization of the original BERT embedding vs $pre^p$ embeddings. Dark areas denote regions of space where the predictor has a high confidence value, whereas lighter areas indicate greater uncertainty. Light blue points indicate female occupations and dark blue points indicate male occupations