To Predict or Not to Predict? Towards reliable uncertainty estimation in the presence of noise

This study evaluates uncertainty estimation methods for multilingual text classification under noisy conditions, finding that while softmax-based approaches struggle in low-resource or domain-shift scenarios, Monte Carlo dropout offers robust calibration and significantly improves performance by enabling the abstention of uncertain predictions.

The Gist

Imagine you are hiring a translator to work on a massive project involving documents in seven different languages. You need them to tell you if a sentence is "simple" (easy to read) or "complex" (hard to read).

Usually, you just ask the translator, "Is this simple or complex?" and they give you an answer. But what if they are guessing? What if they are confident but wrong? In the real world, knowing when a system is unsure is just as important as knowing the answer itself.

This paper is like a rigorous "stress test" for a team of AI translators. The researchers wanted to find the best way to make the AI say, "I'm not sure about this one, please don't count my answer," so that the final results are more reliable.

Here is the breakdown of their findings using some everyday analogies:

1. The Problem: The Overconfident Robot

Imagine an AI that is like a student who memorized the textbook perfectly but has never seen a real exam. When the exam questions are exactly like the textbook (the "in-domain" setting), the student gets an A+ and is very confident.

But, as soon as the exam changes slightly—maybe the font is different, or the questions are about a topic the student didn't study (the "out-of-domain" or noisy setting)—the student starts guessing. The scary part? The student still acts confident. They say, "I'm 99% sure this is a complex sentence!" even when they are wrong.

The researchers asked: How can we make the AI admit when it's guessing?

2. The Tools: Different Ways to Measure "Doubt"

The researchers tested nine different "doubt meters" (Uncertainty Estimation methods) to see which one worked best. Think of these as different ways to check if a student is bluffing:

• The "Softmax" Meter (SR): This is the AI's default confidence score. It's like asking the student, "How sure are you?" on a scale of 1 to 10.
  • Verdict: Great when the test is easy and familiar. But when the test gets weird, the student lies and says "10/10" even when they are clueless.
• The "Monte Carlo Dropout" Meter (SMP/ENT-MC): This is like asking the student the same question 20 times, but every time you ask, you slightly shake their hand or distract them (simulating randomness). If the student gives you 20 different answers, you know they are confused. If they give the same answer 20 times, they are likely right.
  • Verdict: This was the champion. Even when the test got weird or the language was difficult, this method consistently knew when the AI was unsure.
• The "Outlier" Meters (LOF/ISOF): These look at the shape of the data. It's like checking if a sentence looks like a "stranger" compared to the books the AI studied.
  • Verdict: Sometimes they work great at spotting weird sentences, but they are like a moody detective—sometimes they catch the bad guys, sometimes they accuse innocent people. They are too unstable to trust completely.

3. The Big Discovery: "To Predict or Not to Predict?"

The title of the paper asks a philosophical question: Should the AI always try to answer, or should it sometimes say "I don't know"?

The researchers found that saying "I don't know" is a superpower.

• The Experiment: They told the AI to refuse to answer the 10% of sentences it was most unsure about.
• The Result: By simply throwing away the 10% of "risky" guesses, the overall accuracy of the system jumped significantly.
  • Analogy: Imagine a basketball player who misses 30% of their shots. If you tell them, "Don't shoot if you aren't 100% sure," and they only take the easy shots, their scoring percentage skyrockets. The team wins more, even though they took fewer shots.

4. The Takeaway for the Real World

The paper concludes with a practical guide for building better AI:

1. Don't trust the "Confidence Score" blindly: Just because an AI says it's 99% sure doesn't mean it is, especially if the data is noisy or in a new language.
2. The "Shake the Hand" method wins: Using techniques that introduce a little bit of randomness (Monte Carlo Dropout) is the most reliable way to detect when an AI is confused. It's the most robust "lie detector."
3. It's okay to abstain: In critical real-world applications (like medical diagnosis or legal translation), it is better to have the AI say, "I'm not sure, please ask a human," than to have it confidently give a wrong answer.

In summary: This paper teaches us that a smart AI isn't just one that knows all the answers; it's an AI that knows when it doesn't know. By using the right tools to measure doubt and having the courage to skip the hard guesses, we can build AI systems that are much safer and more trustworthy.

Technical Summary

Here is a detailed technical summary of the paper "To Predict or Not to Predict? Towards reliable uncertainty estimation in the presence of noise" by Khallaf and Sharoff.

1. Problem Statement

The paper addresses a critical gap in Natural Language Processing (NLP): the reliability of Uncertainty Estimation (UE) methods in multilingual, noisy, and non-topical environments.

• Context: In real-world applications, classifiers must not only predict but also know when they are likely to be wrong. This is particularly challenging when dealing with:
  • Domain Shift: Models trained on one dataset (e.g., readability assessments) failing on others (e.g., children's encyclopedias or news).
  • Language Shift: Performance degradation when transferring models to low-resource or different languages.
  • Non-Topical Classification: Tasks where the input text does not fit standard topical categories, often leading to "unreasonable" accuracy scores if noise is not handled.
• Core Question: Which UE methods provide the most robust calibration and discriminative power across diverse languages and domain shifts, and how does abstaining from uncertain predictions improve overall system reliability?

2. Methodology

Datasets and Task

• Task: Binary sentence complexity classification (Simple vs. Complex).
• Training Data: Readme++, a multilingual dataset graded by CEFR levels (Simple = B1 or below; Complex = C1/C2).
• Test Data (Evaluation of Robustness):
  • In-Domain: Cross-validation on Readme++.
  • Out-of-Domain (OOD):
    • Vikidia/Wikipedia: Children's content (removes topic bias).
    • Simplext: Manually simplified Spanish news articles.
• Languages: Arabic, Catalan, English, French, Hindi, Russian, Spanish, and Italian.

Models

• Base Classifier: Multilingual BERT (mBERT) fine-tuned with 5-fold cross-validation.
• Note: Larger generative LLMs (GPT, LLaMA) were tested but found to offer no consistent advantage over BERT-sized PLMs for this specific classification task.

Uncertainty Estimation (UE) Methods Evaluated

The study benchmarks 9 UE techniques categorized into three groups:

1. Probabilistic (Output-based):
   • SR (Softmax Response): Max probability from a single pass.
   • SMP (Sampled Max Probability): MC Dropout average of max probabilities.
   • ENT (Entropy): Spread of probability distribution.
   • ENT-MC: Entropy averaged over MC Dropout passes.
   • PV (Probability Variance): Variance of probabilities across MC Dropout runs.
   • BALD: Mutual information between model weights and predictions.
2. Feature-Based (Geometry):
   • MD (Mahalanobis Distance): Distance of test embeddings from training class centroids.
   • LOF (Local Outlier Factor): Local density deviation.
   • ISOF (Isolation Forest): Anomaly detection via tree partitions.
3. Hybrid:
   • HUQ-MD: Combines MD (epistemic) and SR (aleatoric) via rank-based mixing.

Evaluation Metrics

The authors evaluate UE performance across three perspectives:

1. Uncertainty Discrimination: Ability to rank incorrect predictions higher than correct ones (ROC-AUC, AU-PRC).
2. Calibration: Alignment between predicted confidence and actual accuracy (C-Slope, CITL, ECE).
3. Selective Prediction: Improvement in performance when abstaining from uncertain predictions (RC-AUC, N.RC-AUC, E-AUoptRC, Trust Index).

3. Key Contributions

1. Comprehensive Multilingual Benchmark: The first large-scale evaluation of 9 UE methods across 7 languages and 3 datasets, explicitly testing robustness against language and domain shifts.
2. Metric Correlation Analysis: A detailed analysis of how different UE metrics correlate (or fail to correlate) across languages, revealing that high discrimination scores do not always guarantee better selective prediction performance.
3. Practical Insights on Abstention: Demonstrates that abstaining from the top 5–10% most uncertain predictions significantly boosts macro F1 scores, particularly in non-topical settings.
4. Efficiency vs. Robustness Trade-off: Provides runtime overhead data, contrasting computationally cheap methods (SR, ENT) with expensive stochastic methods (MC Dropout).

4. Key Results

Performance Under Domain/Language Shift

• Softmax (SR) Limitations: While SR is competitive in high-resource, in-domain settings (offering good calibration and low computational cost), its reliability declines sharply under domain and language shifts.
• MC Dropout Superiority: SMP and ENT-MC (Monte Carlo Dropout approaches) demonstrate consistently strong performance across all languages and conditions. They offer superior calibration and stable decision thresholds, especially in low-resource scenarios (e.g., Hindi) and OOD settings.
• Feature-Based Instability: Methods like LOF and ISOF often achieve the highest discrimination scores (ROC-AUC) but exhibit high variance and instability across languages. They are less reliable for practical selective prediction.
• Hybrid Performance: HUQ-MD performs well on discrimination but suffers from weaker calibration compared to pure MC Dropout methods.

Selective Prediction (Abstention)

• F1 Improvement: Abstaining from the 10% most uncertain instances in the Readme task increased the macro F1 score from 0.81 to 0.85.
• Threshold Sensitivity: In in-domain settings, low rejection rates (1–15%) yield consistent gains. In OOD settings, gains shrink and become less stable, though MC Dropout methods remain the most robust.
• Error Analysis: The UE methods successfully identified atypical examples (e.g., multi-clause narratives, punctuation-dense structures, or domain shifts like health/tech topics) that the model struggled with.

Efficiency

• Fastest: Feature-based methods (LOF, ISOF) and simple probabilistic methods (SR, ENT) are computationally efficient.
• Slowest: MC Dropout methods (SMP, PV, BALD) incur significant overhead (approx. 21s per fold vs. ~1.3s for others) due to repeated stochastic forward passes.

5. Significance and Conclusion

The paper concludes that there is no single "best" UE method; the choice depends on the operational context:

• In-Domain/High-Resource: Simple, computationally cheap methods like SR and ENT are sufficient and effective.
• Out-of-Domain/Low-Resource/Noisy: MC Dropout-based methods (SMP, ENT-MC) are essential for robust uncertainty estimation, despite their higher computational cost.
• Metric Discrepancy: The study highlights a critical finding: High discrimination metrics (like ROC-AUC) do not always translate to reliable selective prediction gains. Methods that score well on aggregate metrics (like ISOF) may be unstable in practice.

Future Work: The authors propose developing a meta-uncertainty framework that adaptively selects or combines UE metrics based on the specific data context (fold-level and input-level features) to optimize the trade-off between robustness and efficiency in real-world NLP systems.