Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Detecting Errors through Ensembling Prompts (DEEP): An End-to-End LLM Framework for Detecting Factual Errors

Alex Chandler, Devesh Surve, Hui Su

TL;DR

DEEP introduces a full end-to-end framework that detects factual errors in transformer-generated summaries by converting diverse prompts into binary signals and pooling them via ensemble models calibrated for reliable probabilities. It demonstrates that optimizing thresholds on target datasets is crucial for encoder-based factual consistency models, while DEEP achieves state-of-the-art balanced accuracy on AggreFact-XSUM FTSOTA, TofuEval Summary-Level, and HaluEval Summarization without fine-tuning. The approach uses 16 ensembling methods and several calibration strategies (notably Platt Scaling) to produce well-calibrated probabilities, addressing overconfidence issues. The results suggest substantial practical benefits for reliable factuality assessment in real-world, diverse summarization settings, though at higher computational cost.

Abstract

Accurate text summarization is one of the most common and important tasks performed by Large Language Models, where the costs of human review for an entire document may be high, but the costs of errors in summarization may be even greater. We propose Detecting Errors through Ensembling Prompts (DEEP) - an end-to-end large language model framework for detecting factual errors in text summarization. Our framework uses a diverse set of LLM prompts to identify factual inconsistencies, treating their outputs as binary features, which are then fed into ensembling models. We then calibrate the ensembled models to produce empirically accurate probabilities that a text is factually consistent or free of hallucination. We demonstrate that prior models for detecting factual errors in summaries perform significantly worse without optimizing the thresholds on subsets of the evaluated dataset. Our framework achieves state-of-the-art (SOTA) balanced accuracy on the AggreFact-XSUM FTSOTA, TofuEval Summary-Level, and HaluEval Summarization benchmarks in detecting factual errors within transformer-generated text summaries. It does so without any fine-tuning of the language model or reliance on thresholding techniques not available in practical settings.

Detecting Errors through Ensembling Prompts (DEEP): An End-to-End LLM Framework for Detecting Factual Errors

TL;DR

DEEP introduces a full end-to-end framework that detects factual errors in transformer-generated summaries by converting diverse prompts into binary signals and pooling them via ensemble models calibrated for reliable probabilities. It demonstrates that optimizing thresholds on target datasets is crucial for encoder-based factual consistency models, while DEEP achieves state-of-the-art balanced accuracy on AggreFact-XSUM FTSOTA, TofuEval Summary-Level, and HaluEval Summarization without fine-tuning. The approach uses 16 ensembling methods and several calibration strategies (notably Platt Scaling) to produce well-calibrated probabilities, addressing overconfidence issues. The results suggest substantial practical benefits for reliable factuality assessment in real-world, diverse summarization settings, though at higher computational cost.

Abstract

Accurate text summarization is one of the most common and important tasks performed by Large Language Models, where the costs of human review for an entire document may be high, but the costs of errors in summarization may be even greater. We propose Detecting Errors through Ensembling Prompts (DEEP) - an end-to-end large language model framework for detecting factual errors in text summarization. Our framework uses a diverse set of LLM prompts to identify factual inconsistencies, treating their outputs as binary features, which are then fed into ensembling models. We then calibrate the ensembled models to produce empirically accurate probabilities that a text is factually consistent or free of hallucination. We demonstrate that prior models for detecting factual errors in summaries perform significantly worse without optimizing the thresholds on subsets of the evaluated dataset. Our framework achieves state-of-the-art (SOTA) balanced accuracy on the AggreFact-XSUM FTSOTA, TofuEval Summary-Level, and HaluEval Summarization benchmarks in detecting factual errors within transformer-generated text summaries. It does so without any fine-tuning of the language model or reliance on thresholding techniques not available in practical settings.
Paper Structure (17 sections, 1 equation, 4 figures, 6 tables)

This paper contains 17 sections, 1 equation, 4 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Datasets
  3. Methodology
  4. Ensembling Methods
  5. Metrics Used
  6. Calibrating Ensembled Models for Reliable Probability Estimates
  7. Methodology for Prompt Creation
  8. Selection of Prompt Pool
  9. Results
  10. Individual Prompt Results
  11. Ensemble Benchmarking and Comparing to Existing Methods
  12. Calibration Analysis
  13. Statistical Testing
  14. Conclusion
  15. Our Prompts
  16. ...and 2 more sections

Figures (4)

  • Figure 1: Optimal thresholds of factual consistency models when set to maximize balanced accuracy on each test dataset.
  • Figure 2: Factual consistency models' performance varies significantly based on threshold optimization strategy. The bars show balanced accuracy for factual consistency models under three threshold optimization strategies: optimizing the thresholds on the test dataset ("Optimizing on Test"), setting thresholds to the midpoint of each model's score range ("Optimizing at Center"), or optimizing on all datasets except the test set ("Optimizing on Train"), which reflects a realistic scenario of applying the model to unseen data. Numbers above bars quantify the decrease in balanced accuracy when thresholds are optimized on non-test data compared to the test dataset itself, underscoring the difficulty of effectively applying these models to unseen data in practice.
  • Figure 3: Diagram of our end-to-end framework.
  • Figure 4: An example reliability diagram highlighting the difference in reliability between the predicted probabilities before and after calibration. The reliability diagram is separated by positive and negative predictions, highlighting the contrast in model confidences between predicting whether a summary is factually consistent with the source context versus inconsistent. In this visualization, LabelModel was trained with the output of nine LLM Prompts on the non-test datasets and tested on the AggreFact-XSUM FTSOTA test dataset.