Critical scholarship has elevated the problem of gender bias in data sets
used to train virtual assistants (VAs). Most work has focused on explicit
biases in language, especially against women, girls, femme-identifying people,
and genderqueer folk; implicit associations through word embeddings; and
limited models of gender and masculinities, especially toxic masculinities,
conflation of sex and gender, and a sex/gender binary framing of the masculine
as diametric to the feminine. Yet, we must also interrogate how masculinities
are "coded" into language and the assumption of "male" as the linguistic
default: implicit masculine biases. To this end, we examined two natural
language processing (NLP) data sets. We found that when gendered language was
present, so were gender biases and especially masculine biases. Moreover, these
biases related in nuanced ways to the NLP context. We offer a new dictionary
called AVA that covers ambiguous associations between gendered language and the
language of VAs.

研究调查了两个自然语言处理数据集，发现当语言具有性别特征时，特别是具有男性特征时，就会出现性别偏见，为此提出了一个名为 AVA 的新词典，旨在涵盖性别化语言与 VA 语言之间的模糊关联。