Uncertainty Estimation for the Open-Set Text Classification systems
New method predicts when AI text classifiers will fail, improving error detection by up to 365%.
Researchers Leonid Erlygin and Alexey Zaytsev have published a new paper, 'Uncertainty Estimation for the Open-Set Text Classification systems,' introducing a method to make AI text classifiers more trustworthy. The work focuses on Open-Set Text Classification (OSTC), a critical task where a system must either classify text into a known category or reject it as 'unknown.' The team adapted the Holistic Uncertainty Estimation (HolUE) framework to tackle two core sources of error: 'text uncertainty' from poorly formulated user queries and 'gallery uncertainty' from ambiguous or overlapping data in the training set. By quantifying these uncertainties, their model can predict when it is likely to make a mistake, a crucial step for deploying robust AI in high-stakes applications.
To validate their approach, the researchers created a new OSTC benchmark and tested HolUE across diverse datasets including Yahoo Answers for topic classification, DBPedia, PAN for authorship attribution, and CLINC150 for intent recognition. The results are striking: HolUE achieved a 365% improvement in Prediction Rejection Ratio (PRR) on Yahoo Answers (0.79 vs. 0.17 baseline) and a 347% improvement on DBPedia. Even on the more challenging CLINC150 dataset, it delivered a solid 40% gain. The code and protocols are publicly available, providing a practical tool for developers to build AI systems that can confidently say 'I don't know' instead of making a costly, incorrect guess.
- HolUE method identifies 'text uncertainty' from bad queries and 'gallery uncertainty' from ambiguous data to predict AI errors.
- Achieved a 365% improvement in Prediction Rejection Ratio on Yahoo Answers and 347% on DBPedia over baseline methods.
- Publicly released code and a new benchmark enable more robust and trustworthy open-set classification systems for real-world use.
Why It Matters
Enables safer AI deployment by allowing text classifiers to reject uncertain inputs, preventing costly errors in customer service and content moderation.