A Label Distribution for Few-shot In-domain Out-of-Scope Detection

Task oriented dialog systems are widely used in everyday scenarios. Intent classification is the essence of task-oriented dialog systems automatically classifying user inquiries into different categories based on their diverse intentions. Current state-of-the-art methods are efficient for most out-of-scope detection tasks, whilst these methods are insufficient capable to identify out-of-scope queries which are similar to in-scope queries. Moreover, there is a scarcity of well-defined queries for known classes in real applications. Thus, few-shot intent classification is another key aspect within our task. In this paper, we first identify in-domain out-of-scope detection where most of out-of-scope queries are similar to in-scope queries. We provide a label distribution with Local Outlier Factor to efficiently tackle few-shot in-domain out-of-scope detection. Furthermore, experiment results are provided to show the effectiveness of our approach on standard benchmark datasets.