Whose Language Counts as High Quality? Measuring Language Ideologies in Text Data Selection

Language models increasingly rely on massive 001 web dumps for diverse text data. However, 002 these sources are rife with undesirable content. 003 As such, resources like Wikipedia, books, and 004 news often serve as anchors for automatically 005 selecting web text most suitable for language 006 modeling, a process typically referred to as 007 quality filtering. Using a new dataset of U.S. 008 high school newspaper articles—written by stu009 dents from across the country—we investigate 010 whose language is preferred by the quality fil011 ter used for GPT-3. We find that newspapers 012 from larger schools, located in wealthier, edu013 cated, and urban ZIP codes are more likely to 014 be classified as high quality. We then demon015 strate that the filter’s measurement of quality 016 is unaligned with other sensible metrics, such 017 as factuality or literary acclaim. We argue that 018 privileging any corpus as high quality entails a 019 language ideology, and more care is needed to 020 construct training corpora for language models, 021 with better transparency and justification for 022 the inclusion or exclusion of various texts. 023