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

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 2562 - 2580 December 7-11, 2022 ©2022 Association for Computational Linguistics Whose Language Counts as High Quality? Measuring Language Ideologies in Text Data Selection Suchin Gururangan† Dallas Card♢ Sarah K. Dreier♡ Emily K. Gade♣ Leroy Z. Wang† Zeyu Wang† Luke Zettlemoyer† Noah A. Smith†♠ †University of Washington ♢University of Michigan ♡University of New Mexico ♣Emory University ♠Allen Institute for AI {sg01,zwan4,lsz,nasmith}@cs.washington.edu dalc@umich.edu skdreier@unm.edu emily.gade@emory.edu lryw@uw.edu Abstract Language models increasingly rely on massive web crawls for diverse text data. However, these sources are rife with undesirable content. As such, resources like Wikipedia, books, and news 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 stu- dents 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 zones (ZIP codes) are more likely to be classified as high quality. We also show that this quality measurement is unaligned with other sensible metrics, such as factuality or lit- erary acclaim. We argue that privileging any corpus as high quality entails a language ideol- ogy, and more care is needed to construct train- ing corpora for language models, with better transparency and justification for the inclusion or exclusion of various texts. 1 Introduction The language models central to modern NLP are trained on large Internet corpora, typically gathered from community resources (e.g., Wikipedia; Liu et al. 2019) or web crawls (e.g., WebText, Common Crawl; Radford et al. 2019, Brown et al. 2020). The selection of texts impacts every research or deployed NLP system that builds on these models. Yet there is rarely any clear justification given for why various texts were included. Web dumps like Common Crawl offer the promise of more diverse text than what is avail- able in curated resources. However, much of the web consists of frequently replicated boilerplate (e.g., privacy policies), code (e.g., HTML and Javascript), pornography, hate speech, and more. Automated approaches, typically referred to as quality filters, are often applied in an effort to re- move this undesirable content from training data.1 These filters include code removers (Gao et al., 2020), heuristics (Rae et al., 2021), stopwords (Raf- fel et al., 2020), and classifiers (Brown et al., 2020; Wenzek et al., 2020). Although quality filtering is often treated as a relatively neutral preprocessing step, it necessarily implies a value judgment: which data is assumed to be of sufficiently high quality to be included in the training corpus? More concretely, when a quality filter is a classifier trained on instances assumed to be of high (and low) quality, the selection of those examples will impact the language model and any downstream technology that uses it. Many filters use Wikipedia, books, and newswire to represent high quality text. But what texts are excluded as …

Content truncated to 500 words