Disentangling Confidence Score Distribution for Out-of-Domain Intent Detection with Energy-Based Learning
This addresses a critical issue for task-oriented dialog systems by enhancing reliability in handling unknown user queries, though it is incremental as it builds on existing energy-based approaches.
The paper tackles the problem of detecting out-of-domain intents in dialog systems, where traditional methods suffer from overconfidence, and proposes an energy-based score function that achieves improved OOD detection by disentangling confidence score distributions, with experimental results showing significant gains over baselines.
Detecting Out-of-Domain (OOD) or unknown intents from user queries is essential in a task-oriented dialog system. Traditional softmax-based confidence scores are susceptible to the overconfidence issue. In this paper, we propose a simple but strong energy-based score function to detect OOD where the energy scores of OOD samples are higher than IND samples. Further, given a small set of labeled OOD samples, we introduce an energy-based margin objective for supervised OOD detection to explicitly distinguish OOD samples from INDs. Comprehensive experiments and analysis prove our method helps disentangle confidence score distributions of IND and OOD data.\footnote{Our code is available at \url{https://github.com/pris-nlp/EMNLP2022-energy_for_OOD/}.}