3 years ago
Minority Voices ‘Filtered’ Out of Google Natural Language Processing Models

Originally published in United AI, Sept 24, 2021.

According to new research, one of the largest Natural Language Processing (NLP) datasets available has been extensively ‘filtered’ to remove black and Hispanic authors, as well as material related to gay and lesbian identities, and source data that deals with a number of other marginal or minority identities.

The dataset was used to train Google’s Switch Transformer and T5 model, and was curated by Google AI itself.

The report asserts that the Colossal Clean Crawled Corpus (‘C4’) dataset, which contains 156 billion tokens scraped from more than 365 million internet domains, and is a subset of the massive Common Crawl scraped database, has been extensively (algorithmically) filtered to exclude ‘offensive’ and ‘toxic’ content, and that the filters used to distill C4 have effectively targeted content and discussion from minority groups.

The report states:

‘Our examination of the excluded data suggests that documents associated with Black and Hispanic authors and documents mentioning sexual orientations are significantly more likely to be excluded by C4.EN’s blocklist filtering, and that many excluded documents contained non-offensive or non-sexual content (e.g., legislative discussions of same-sex marriage, scientific and medical content).’

To continue reading this article, click here.