Few-Shot Out of Domain Intent Detection with Covariance Corrected Mahalanobis Distance
A paper proposes a covariance corrected Mahalanobis distance for detecting out-of-domain intents in few-shot settings, addressing limitations of prior work by Podolskiy et al. (2021).
The paper analyzes why the Mahalanobis distance method by Podolskiy et al. (2021) underperforms in few-shot out-of-domain intent detection and proposes a covariance corrected Mahalanobis distance to improve performance.
The proposed method corrects the covariance estimation in the Mahalanobis distance to better handle few-shot scenarios, potentially improving OOD intent detection accuracy.
Improved few-shot OOD intent detection could enhance robustness of conversational agents in production with limited labeled data.
Potential to reduce false positives for unknown intents in chatbots and voice assistants, improving user experience and reducing manual review.
Further validation on benchmark datasets and comparison with other few-shot OOD methods would be needed to assess practical impact.