A Causal Graph for Learning Gender Debiased Word Embedding

Word embedding learning is a powerful technique to represent words’ rich semantics as low-dimensional vectors, but it may encode harmful social biases. Such biases can leave negative impacts on downstream applications that rely on word embeddings, e.g., text classification and sentiment analysis. In this work, we design a novel approach to debias word embeddings using a causal graph that separates gender and semantic information within words. Our method is theoretically interpretable and can effectively alleviate the issue of gender bias from words in real-world corpora without destructing their semantic meanings. We perform a comprehensive comparison for our proposed approach with existing state-of-the-art techniques using a range of evaluation metrics and tasks, and demonstrate that our method can not only mitigate bias mixed in word embeddings, but also improve their performance on downstream tasks.