Performance and Risk Trade-offs for Multi-word Text Prediction at Scale

Findings of the Association for Computational Linguistics: EACL 2023, pages 2226–2242 May 2-6, 2023 ©2023 Association for Computational Linguistics Performance and Risk Trade-offs for Multi-word Text Prediction at Scale Warning: The paper contains examples which the reader might find offensive. Aniket Vashishtha S Sai Krishna Prasad Payal Bajaj Vishrav Chaudhary Kate Cook Sandipan Dandapat Sunayana Sitaram Monojit Choudhury Microsoft Corporation {t-aniketva,sai.krishna,payal.bajaj,vchaudhary, katherine.cook,sadandap,sunayana.sitaram,monojitc}@microsoft.com Abstract Large Language Models such as GPT-3 are well-suited for text prediction tasks, which can help and delight users during text composition. LLMs are known to generate ethically inappro- priate predictions even for seemingly innocu- ous contexts. Toxicity detection followed by filtering is a common strategy for mitigating the harm from such predictions. However, as we shall argue in this paper, in the context of text prediction, it is not sufficient to detect and filter toxic content. One also needs to ensure factual correctness and group-level fairness of the predictions; failing to do so can make the system ineffective and nonsensical at best, and unfair and detrimental to the users at worst. We discuss the gaps and challenges of toxic- ity detection approaches – from blocklist-based approaches to sophisticated state-of-the-art neu- ral classifiers – by evaluating them on the text prediction task for English against a manually crafted CheckList of harms targeted at different groups and different levels of severity. 1 Introduction Large Language Models (LLMs) are powerful, yet known to generate potentially risky, harmful, of- fensive texts (Bender et al., 2021; Weidinger et al., 2021), even when the context is seemingly innocu- ous (Gehman et al., 2020). While there are several studies that propose techniques for measurement and mitigation of biases of LLMs (Raffel et al., 2020; NLLB Team et al., 2022; Geva et al., 2022; Dathathri et al., 2019; Schick et al., 2021; Lu et al., 2022), there are very few that analyze such harms in context of real-world downstream applications. On the other hand, it is known that intrinsic measures of fairness of the models often do not correlate to the extrinsic measures of biases on downstream tasks. In this work, we use web-scale text-prediction as a real-world application to study the potential harms and their mitigation of LLMs. Text predic- tion is a commonly used writing assistance technol- ogy which provides real-time, context-dependent word and phrase completions as a user is typing (Dai et al., 2019; Garay-Vitoria and Abascal, 2006; Hard et al., 2018). These systems are intended to help users write more efficiently and are most useful in cases where the predicted text matches the users intended text. Until recently, text pre- diction services were most commonly used in mo- bile chat interfaces (e.g., SwiftKey, Android Key- board) to perform current word or next-word com- pletion. Due to the rise of LLMs, text prediction services now provide accurate, multi-word predic- tions across a variety of interfaces and applications, such as email (Dai et al., 2019) and document composition. Text prediction can affect the process and the product of writing (Bhat et al., 2021; …

Content truncated to 500 words