D2U: Distance-to-Uniform Learning for Out-of-Scope Detection

Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2093 - 2108 July 10-15, 2022 ©2022 Association for Computational Linguistics D2U: Distance-to-Uniform Learning for Out-of-Scope Detection Eyup Halit Yilmaz and Cagri Toraman Aselsan Research Center Ankara, Turkey {ehyilmaz,ctoraman}@aselsan.com.tr Abstract Supervised training with cross-entropy loss implicitly forces models to produce probabil- ity distributions that follow a discrete delta dis- tribution. Model predictions in test time are expected to be similar to delta distributions if the classifier determines the class of an in- put correctly. However, the shape of the pre- dicted probability distribution can become sim- ilar to the uniform distribution when the model cannot infer properly. We exploit this obser- vation for detecting out-of-scope (OOS) utter- ances in conversational systems. Specifically, we propose a zero-shot post-processing step, called Distance-to-Uniform (D2U), exploiting not only the classification confidence score, but the shape of the entire output distribution. We later combine it with a learning procedure that uses D2U for loss calculation in the su- pervised setup. We conduct experiments using six publicly available datasets. Experimental results show that the performance of OOS de- tection is improved with our post-processing when there is no OOS training data, as well as with D2U learning procedure when OOS train- ing data is available. 1 Introduction Automated conversational systems have recently received attention from the research community (Dopierre et al., 2021; Mehri et al., 2020; Qin et al., 2021). In applications such as voice assistants, Spo- ken Language Understanding (Young et al., 2013) aims to extract meaning from the user inputs, called utterances, in order to process and execute desired functionalities. The task of Intent Detection, or Intent Classification, aims to classify user utter- ance into a set of system-identifiable intents. How- ever, supervised training of such systems can only cover a restricted set of classes, i.e. in-scope (INS) classes. To enhance user experience, the task of Out-of-Scope (OOS) detection (Lin and Xu, 2019a; Xu et al., 2021; Zhan et al., 2021; Shen et al., 2021) RateBook PlayMusic GetWeather AddToPlaylist BookRestaurant SearchCreativeWork SearchScreeningEvent 0.0 0.2 0.4 0.6 0.8 1.0 Probabilty Scores INS Utterance RateBook PlayMusic GetWeather AddToPlaylist BookRestaurant SearchCreativeWork SearchScreeningEvent 0.0 0.2 0.4 0.6 0.8 1.0 OOS Utterance Figure 1: Sample output distributions of an INS classi- fier predicting the intent of an INS and OOS utterance. Since OOS utterances do not belong to any intent, the prediction gets closer to the uniform distribution. distinguishes INS utterances from those that do not belong to the scope of the classifier with dedicated model architectures and loss functions. Existing methods in OOS detection utilize the classifier confidence score for a given utterance with thresholding to classify highly confident pre- dictions as INS and lower confidence predictions as OOS. However, softmax classifiers suffer from overconfident predictions for OOS data (Hendrycks and Gimpel, 2017), which makes it difficult to ac- curately determine a threshold value. Confidence loss (Lee et al., 2018) mitigates this by calculat- ing the KL Divergence …

Content truncated to 500 words