Prompting PaLM for Translation: Assessing Strategies and Performance

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 1533–1545 July 9-14, 2023 ©2023 Association for Computational Linguistics On Prefix-tuning for Lightweight Out-of-distribution Detection Yawen Ouyang Yongchang Cao Yuan Gao Zhen Wu Jianbing Zhang Xinyu Dai National Key Laboratory for Novel Software Technology, Nanjing University, China Collaborative Innovation Center of Novel Software Technology and Industrialization, China {ouyangyw, caoyc, gaoy}@smail.nju.edu.cn {wuz, zjb, daixinyu}@nju.edu.cn Abstract Out-of-distribution (OOD) detection, a funda- mental task vexing real-world applications, has attracted growing attention in the NLP commu- nity. Recently fine-tuning based methods have made promising progress. However, it could be costly to store fine-tuned models for each scenario. In this paper, we depart from the clas- sic fine-tuning based OOD detection toward a parameter-efficient alternative, and propose an unsupervised prefix-tuning based OOD detec- tion framework termed PTO. Additionally, to take advantage of optional training data labels and targeted OOD data, two practical exten- sions of PTO are further proposed. Overall, PTO and its extensions offer several key advan- tages of being lightweight, easy-to-reproduce, and theoretically justified. Experimental results show that our methods perform comparably to, even better than, existing fine-tuning based OOD detection approaches under a wide range of metrics, detection settings, and OOD types. 1 Introduction Detecting out-of-distribution (OOD) inputs is cru- cial for real-world machine learning systems de- ployed in the wild (Hendrycks and Gimpel, 2017). For example, for a task-oriented dialogue system designed for particular domains, it can be challeng- ing to ensure that the system is only exposed to utterances from the same distribution as the train- ing utterances, i.e., in-distribution (ID) utterances. Therefore, it would be desirable for the system to detect OOD utterances and return safe responses. Pretrained language models (PLMs) have been a de facto choice for OOD detection in the NLP community, and many fine-tuning based methods have achieved promising results (Arora et al., 2021; Podolskiy et al., 2021; Lang et al., 2022). Despite being effective, these methods require storing fine- tuned models for each scenario, which could be prohibitively expensive. This begs the following question: Can we achieve effective OOD detec- tion in a parameter-efficient way, i.e., keep PLM parameters frozen? To achieve this goal, an unsupervised Prefix- Tuning based OOD detection framework (PTO) is proposed in this paper. The key idea of PTO is intuitive: an in-distribution specific prefix, op- timized with the training data via maximum like- lihood, could steer PLMs to assign higher likeli- hoods to ID samples than PLMs without the pre- fix, while OOD samples should be assigned lower likelihood. Thus we propose to use the likelihood change triggered by the prefix to detect OOD — samples whose improvement is not obvious (e.g., less than a predefined threshold). Note that the training process of PTO does not involve the sam- ple labels, expanding its application to situations where obtaining labeled data is cost-prohibitive. Going beyond the unsupervised setting, we ex- tend our framework to fully leverage optional super- vised data. Specifically, we design two …

Content truncated to 500 words