Revisiting Mahalanobis Distance for Transformer-Based Out-of-Domain Detection

Abstract

Real-life applications, heavily relying on machine learning, such as dialog systems, demand for out-of-domain detection methods. Intent classification models should be equipped with a mechanism to distinguish seen intents from unseen ones so that the dialog agent is capable of rejecting the latter and avoiding undesired behavior. However, despite increasing attention paid to the task, the best practices for out-of-domain intent detection have not yet been fully established. This paper conducts a thorough comparison of out-of-domain intent detection methods. We prioritize the methods, not requiring access to out-of-domain data during training, gathering of which is extremely time- and labor-consuming due to lexical and stylistic variation of user utterances. We evaluate multiple contextual encoders and methods, proven to be efficient, on three common datasets for intent classification, expanded with out-of-domain utterances. Our main findings show that fine-tuning Transformer-based encoders on in-domain data leads to superior results. Mahalanobis distance, together with utterance representations, derived from Transformer-based encoders, outperform other methods by a wide margin(1-5% in terms of AUROC) and establish new state-of-the-art results for all datasets. The broader analysis shows that the reason for success lies in the fact that the fine-tuned Transformer is capable of constructing homogeneous representations of in-domain utterances, revealing geometrical disparity to out of domain utterances. In turn, the Mahalanobis distance captures this disparity easily.

Fine-Tuning Deteriorates General Textual Out-of-Distribution Detection by Distorting Task-Agnostic Features

“…Nonetheless, almost all of the current approaches are assessed under certain assumptions about the type of OOD texts. One line of works creates in-distribution (ID) and OOD pairs from arbitrary datasets for different tasks (Hendrycks et al, 2020), while another line assumes that OOD data belong to classes in the ID task but unseen during training, e.g., in intent recognition (Podolskiy et al, 2021). Arora et al (2021) reveal the inconsistency among the evaluation protocols and category the distribution shifts to non-semantic shifts (NSS) and semantic shifts (SS), but a thorough comparison of existing methods in different settings is missing as later works either focus on either detecting NSS (Duan et al, 2022) or SS (Zhou et al, 2022).…”

“…Notably, the detec-tors based on the pre-trained features, e.g., the Mahalanobis distance detector MD pre (Xu et al, 2021), excel at detecting non-semantic shifts but fail in detecting semantic shifts. In contrast, when the PLM is fine-tuned on annotated ID data, the detectors based on fine-tuned features, e.g., MD ft (Podolskiy et al, 2021), perform well in the SS scenario but disastrously fail in the NSS setting. These observations uncover an intriguing trade-off: fine-tuning contributes to the detection of semantic shifts but impairs the detection of non-semantic shifts.…”

“…As for the density-based branch, Gangal et al (2020) and Arora et al (2021) apply the idea to textual OOD detection by leveraging language models such as LSTM (Hochreiter and Schmidhuber, 1997) and GPT-2 (Radford et al, 2019). Regarding the distance-based methods, Podolskiy et al (2021) revisit the Mahalanobis distance-based detector (Lee et al, 2018) for textual OOD detection based on fine-tuned PLMs and achieve performance gains over confidence-based methods, which is then further improved by introducing contrastive regularization , utilizing nearestneighbor distance (Zhou et al, 2022), and leveraging intermediate features (Chen et al, 2022).…”

“…(2) the semantic shift (SS) setting (Podolskiy et al, 2021;Zhou et al, 2022), where OOD data are composed of unseen classes belonging to the ID task, e.g., new classes in intent classification;…”

“…However, fine-tuned pre-trained language models (PLMs) notoriously suffer from over-confident predictions on out-ofdistribution (OOD) inputs (Hendrycks et al, 2020). As this issue threats the reliability of NLP models deployed in the open world, textual OOD detection has attracted great attention recently (Podolskiy et al, 2021;Zhou et al, , 2022Duan et al,…”