Introduction
Modern deep learning models in Natural Language Processing (NLP) have achieved remarkable accuracy across a wide range of tasks. However, they often exhibit a critical flaw: overconfidence. As models grow more complex, they frequently assign high probability scores to incorrect predictions, making them "unaware of what they don't know." In high-stakes domains such as medicine or law, an incorrect prediction can be far more damaging than a simple refusal to answer.
This paper investigates selective prediction (or abstention), a framework that allows a model to withhold a prediction when its confidence falls below a certain threshold. By enabling models to say "I don't know," we can improve reliability and facilitate human-in-the-loop systems or multi-model cascades.
Methodology: Selective Classification and Error Regularization
A selective classifier is defined as a pair $(f, g)$, where $f$ is the primary classification model and $g$ is a selection function. The selection function $g(x)$ determines whether to output the model's prediction $f(x)$ or an abstention symbol ($\perp$).
The authors propose a novel Error Regularization loss ($L_{reg}$) to improve the model's ability to rank its own confidence. Unlike standard cross-entropy loss, which only focuses on correctness, the error regularizer penalizes the model when it is more confident about a "difficult" example than an "easy" one. Difficulty is quantified using two metrics:
- Current Iteration Error: Whether the model is currently incorrect on the example.
- History Record Error: The frequency of incorrect predictions for a specific example throughout the training process, identifying "forgettable" or inherently ambiguous instances.
By minimizing the combined loss, the model learns a Softmax Response (SR) that more accurately reflects its true likelihood of being correct.
Experimental Evaluation
The proposed methods were evaluated on several GLUE benchmark tasks, including MRPC (paraphrase detection), QNLI (natural language inference), and MNLI (multi-genre NLI). To test the models' ability to handle out-of-distribution data, the researchers also developed "no-answer" benchmarks (bMNLI and bSST-5), where certain classes were excluded from training but present during testing.
Key Results and Insights
The experimental results demonstrate that error regularization consistently improves selective performance.
- Performance Gains: The proposed regularization reduced the Area Under the Risk-Coverage Curve (AURCC) and the Reversed Pair Proportion (RPP) by approximately 10% across various Transformer-based models.
- Softmax Response vs. MC-Dropout: While Monte-Carlo (MC) Dropout is a standard technique for uncertainty estimation in computer vision, the authors found it to be less effective and more computationally expensive than simple Softmax Response (SR) in NLP contexts.
- Efficiency via Cascades: Improved confidence estimation enables the creation of "classifier cascades." In these systems, a small, efficient model (like an LSTM) handles easy examples, while a larger, more accurate model (like BERT) is only invoked for difficult cases. Regularized models achieved a superior accuracy-efficiency frontier in these scenarios.
Conclusion
This work highlights that for reliable NLP systems, ranking confidence is often more important than absolute calibration. By incorporating error regularization, models can be trained to better distinguish between what they have learned and what remains ambiguous. Future research will explore extending these techniques to sequence labeling and generative tasks, as well as incorporating varying costs for different types of errors and abstentions.
Summary updated