Unsupervised Out-of-Domain Detection via Pre-trained Transformers

Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 1052–1061 August 1–6, 2021. ©2021 Association for Computational Linguistics 1052 Unsupervised Out-of-Domain Detection via Pre-trained Transformers Keyang Xu1, Tongzheng Ren2, Shikun Zhang3, Yihao Feng2, Caiming Xiong4 1 Columbia University 2 University of Texas at Austin 3 Carnegie Mellon University 4 Salesforce Research kx2155@columbia.edu {tzren, yihao}@cs.utexas.edushikunz@cs.cmu.edu cxiong@salesforce.com Abstract Deployed real-world machine learning appli- cations are often subject to uncontrolled and even potentially malicious inputs. Those out- of-domain inputs can lead to unpredictable outputs and sometimes catastrophic safety is- sues. Prior studies on out-of-domain detec- tion require in-domain task labels and are lim- ited to supervised classification scenarios. Our work tackles the problem of detecting out-of- domain samples with only unsupervised in- domain data. We utilize the latent represen- tations of pre-trained transformers and pro- pose a simple yet effective method to trans- form features across all layers to construct out- of-domain detectors efficiently. Two domain- specific fine-tuning approaches are further pro- posed to boost detection accuracy. Our em- pirical evaluations of related methods on two datasets validate that our method greatly im- proves out-of-domain detection ability in a more general scenario.1 1 Introduction Deep neural networks, despite achieving good per- formance on many challenging tasks, can make overconfident predictions for completely irrelevant and out-of-domain (OOD) inputs, leading to sig- nificant AI safety issues (Hendrycks and Gimpel, 2017). Detecting out-of-domain inputs is a funda- mental task for trustworthy AI applications in real- world use cases, because those applications are of- ten subject to ill-defined queries or even potentially malicious inputs. Prior work on out-of-domain de- tection (e.g., Hendrycks and Gimpel, 2017; Lee et al., 2018; Liang et al., 2018; Hendrycks et al., 2019, 2020; Xu et al., 2020) mostly requires in- domain task labels, limiting its usage to super- vised classification. However, deployed applica- 1Code is available at https://github.com/rivercold/ BERT-unsupervised-OOD. tions rarely receive controlled inputs and are sus- ceptible to an ever-evolving set of user inputs that are scarcely labeled. For example, for many non- classification tasks, such as summarization or topic modeling, there are no available classifiers or task labels, which limits the practical usage of recently proposed out-of-domain detection methods. There- fore, it is natural to ask the following question: Can we detect out-of-domain samples using only unsupervised data without any in-domain labels? We regard the out-of-domain detection problem as checking whether the given test samples are drawn from the same distribution that generates the in-domain samples, which requires a weaker assumption than prior work (e.g., Lee et al., 2018; Hendrycks et al., 2020). We suppose that there are only in-domain samples, which allows us to under- stand the properties of data itself regardless of tasks. Therefore, methods developed for this problem are more applicable than task-specific ones and can be further adapted to tasks where no classification la- bels are present, such as active learning or transfer learning. To solve the problem, we utilize the latent …

Content truncated to 500 words