Research Repository

Word Embeddings via Causal Inference: Gender Bias Reducing and Semantic Information Preserving

Ding, Lei and Yu, Dengdeng and Xie, Jinhan and Guo, Wenxing and Hu, Shenggang and Liu, Meichen and Kong, Linglong and Dai, Hongsheng and Bao, Yanchun and Jiang, Bei (2021) Word Embeddings via Causal Inference: Gender Bias Reducing and Semantic Information Preserving. In: Thirty-Sixth AAAI Conference on Artificial Intelligence (AAAI-22), 2022-02-22 - 2022-03-01, Vancouver. (In Press)

[img]
Preview
Text
word_debias.pdf - Accepted Version

Download (770kB) | Preview

Abstract

With widening deployments of natural language processing (NLP) in daily life, inherited social biases from NLP models have become more severe and problematic. Previous studies have shown that word embeddings trained on human-generated corpora have strong gender biases that can produce discriminative results in downstream tasks. Previous debiasing methods focus mainly on modeling bias and only implicitly consider semantic information while completely overlooking the complex underlying causal structure among bias and semantic components. To address these issues, we propose a novel methodology that leverages a causal inference framework to effectively remove gender bias. The proposed method allows us to construct and analyze the complex causal mechanisms facilitating gender information flow while retaining oracle semantic information within word embeddings. Our comprehensive experiments show that the proposed method achieves state-of-the-art results in gender-debiasing tasks. In addition, our methods yield better performance in word similarity evaluation and various extrinsic downstream NLP tasks.

Item Type: Conference or Workshop Item (Paper)
Additional Information: Published proceedings: _not provided_
Divisions: Faculty of Science and Health
Faculty of Science and Health > Mathematical Sciences, Department of
SWORD Depositor: Elements
Depositing User: Elements
Date Deposited: 18 Feb 2022 09:13
Last Modified: 18 Feb 2022 09:13
URI: http://repository.essex.ac.uk/id/eprint/31799

Actions (login required)

View Item View Item