Combining multiple source embeddings to create meta-embeddings is considered effective to obtain more accurate embeddings. Different methods have been proposed to develop meta-embeddings from a given set of source embeddings. However, the source embeddings can contain unfair gender bias, and the bias in the combination of multiple embeddings and debiasing it effectively have not been studied yet. In this paper, we investigate the bias in three types of meta-embeddings: (1) Multi-Source No-Debiasing: meta-embedding from multiple source embeddings without any debiasing. The experimental results show that meta-embedding amplifies the gender bias compared to those of input source embeddings; (2) Multi-Source Single-Debiasing: meta-embedding from multiple source embeddings debiased by a single method and it can be created in three ways: debiasing each source embedding, debiasing the learned meta-embeddings, and debiasing both source embeddings and meta-embeddings. The results show that debiasing both is the best in two out of three bias evaluation benchmarks; (3) Single-Source Multi-Debiasing: meta-embedding from the same source embedding debiased by different methods. It performed more effectively than its source embeddings debiased with a single method in all three bias evaluation benchmarks.

本文研究了从给定的源嵌入集合中开发元嵌入的不同方法，以及它们中的性别偏见如何影响元嵌入，并通过实验证明，debiasing不仅需要源代码，还需要它们的元嵌入。作者提出了一种基于元嵌入学习的新型去偏方法，其中使用多种去偏方法对单个源嵌入进行处理，然后创建一个单一的无偏元嵌入。

元嵌入中的性别偏见