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

TL;DR

This paper examines how language models' quality filters favor certain socio-economic and geographic language varieties, revealing biases and the need for more transparent data selection practices.

Contribution

It introduces a new dataset and analysis showing biases in quality filtering, highlighting the influence of language ideologies on training data for language models.

Findings

Larger, wealthier, urban schools' newspapers are more likely to be classified as high quality.

The quality filter's measurement does not align with factuality or literary merit.

Biases in data selection reflect underlying language ideologies.

Abstract

Language models increasingly rely on massive web dumps for diverse text data. However, these sources are rife with undesirable content. As such, resources like Wikipedia, books, and newswire often serve as anchors for automatically selecting web text most suitable for language modeling, a process typically referred to as quality filtering. Using a new dataset of U.S. high school newspaper articles -- written by students from across the country -- we investigate whose language is preferred by the quality filter used for GPT-3. We find that newspapers from larger schools, located in wealthier, educated, and urban ZIP codes are more likely to be classified as high quality. We then demonstrate that the filter's measurement of quality is unaligned with other sensible metrics, such as factuality or literary acclaim. We argue that privileging any corpus as high quality entails a language ideology, and more care is needed to construct training corpora for language models, with better transparency and justification for the inclusion or exclusion of various texts.