EMNLP 2024
Detecting factual errors with DEEP
Ensembling prompt judgments to detect factual inconsistencies in generated summaries.
What kind of error are we looking for?
DEEP combines different evaluation perspectives. The terminology map explores the definitions behind related error types. It is a later research project, not the taxonomy used to train DEEP.
Project description
From the original project page. Results and time references describe that version of the work.
Overview
In this paper accepted at EMNLP 2024, 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.
From error definitions to judgments
Different prompts examine factual support, entities, quantities and other error criteria in the same source and summary. Each prompt is evaluated by an LLM. Its binary output becomes an input to the ensemble, whose prediction is then calibrated. The diagram shows illustrative perspectives, not three fixed prompts or three separate models.
The question
DEEP stands for Detecting Errors through Ensembling Prompts. It studies how to assess a generated summary against its source when individual LLM judgments disagree.
The approach
Different prompts produce binary judgments. Ensembling models combine these features, followed by probability calibration. The repository focuses on the ensemble methods and calibration.
The finding
The paper reports leading balanced accuracy on three summarization benchmarks at publication: AggreFact-XSUM FTSOTA, TofuEval Summary-Level, and HaluEval Summarization. These are benchmark results, not a guarantee for every domain.
Why evaluation design matters
Choosing a threshold using the test data can overstate practical performance. DEEP examines what happens when that information is unavailable at deployment.