[Paper Review] Confidence Modeling for Neural Semantic Parsing

Li Dong, Chris Quirk+1 more|arXiv (Cornell University)|May 11, 2018

Topic Modeling27 references3 citations

TL;DR

This paper proposes a confidence modeling framework for neural semantic parsers that identifies uncertainty through three sources—model, data, and input uncertainty—using custom metrics and a regression model to predict confidence scores. It outperforms posterior probability baselines and enables interpretable uncertainty attribution via backpropagation-based saliency, improving both accuracy and model interpretability on IFTTT and Django datasets.

ABSTRACT

In this work we focus on confidence modeling for neural semantic parsers which are built upon sequence-to-sequence models. We outline three major causes of uncertainty, and design various metrics to quantify these factors. These metrics are then used to estimate confidence scores that indicate whether model predictions are likely to be correct. Beyond confidence estimation, we identify which parts of the input contribute to uncertain predictions allowing users to interpret their model, and verify or refine its input. Experimental results show that our confidence model significantly outperforms a widely used method that relies on posterior probability, and improves the quality of interpretation compared to simply relying on attention scores.

Motivation & Objective

• To address the lack of interpretability and confidence estimation in neural semantic parsers, which often act as black boxes.
• To identify and quantify three sources of uncertainty: model, data, and input uncertainty in sequence-to-sequence semantic parsing models.
• To develop a confidence estimation framework that improves upon posterior probability as a confidence indicator in neural networks.
• To enable model interpretation by identifying which input tokens contribute most to prediction uncertainty.
• To support practical deployment by enabling thresholding confidence scores for coverage-accuracy trade-offs.

Proposed method

• Categorize uncertainty into three types: model uncertainty (e.g., overfitting), data uncertainty (e.g., noisy training data), and input uncertainty (e.g., ambiguous or rare inputs).
• Design specific metrics for each uncertainty type, such as prediction variance, data distribution distance, and input token saliency.
• Train a regression model on held-out data using these metrics as features to predict F1 scores, which are then used as confidence scores at test time.
• Apply uncertainty backpropagation to compute gradient-based saliency scores, identifying input tokens that most influence prediction uncertainty.
• Use Gaussian noise perturbations during inference to estimate uncertainty contributions via weight perturbation and gradient computation.
• Combine confidence scores with attention mechanisms to improve interpretability, avoiding reliance on attention alone.

Experimental results

Research questions

• RQ1 Can confidence estimation in neural semantic parsers be improved beyond relying on posterior probabilities?
• RQ2 To what extent can uncertainty be decomposed into model, data, and input components for better interpretability?
• RQ3 Can backpropagation-based saliency provide more accurate and meaningful uncertainty attribution than attention scores?
• RQ4 Does thresholding confidence scores lead to a better trade-off between coverage and accuracy in semantic parsing?
• RQ5 Can the proposed framework be applied to diverse sequence-to-sequence architectures without retraining?

Key findings

• The proposed confidence model significantly outperforms the posterior probability baseline in identifying correct predictions, especially in low-confidence regimes.
• On the IFTTT and Django datasets, the method achieves an overlap@4 of 0.791 and 0.788 respectively, indicating strong alignment with gold-standard uncertain tokens.
• Uncertainty backpropagation identifies input tokens contributing to uncertainty with higher fidelity than attention mechanisms, as shown in qualitative examples.
• The model improves interpretation quality by highlighting function arguments (e.g., URLs, message content) and ambiguous inputs as sources of uncertainty.
• Thresholding confidence scores enables a favorable trade-off between coverage and accuracy, supporting practical deployment in dialogue systems.
• The framework is architecture-agnostic and does not interfere with model training, making it broadly applicable to various sequence-to-sequence models.