text
stringlengths
0
89.3k
an endtoend large language model frame
work for detecting factual errors in text summa
rization Our framework uses a diverse set of
LLM prompts to identify factual inconsisten
cies 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 fac
tual errors in summaries perform significantly
worse without optimizing the thresholds on sub
sets of the evaluated dataset Our framework
achieves stateoftheart SOTA balanced ac
curacy on the AggreFactXSUM FTSOTA To
fuEval SummaryLevel and HaluEval Summa
rization benchmarks in detecting factual errors
within transformergenerated text summaries
It does so without any finetuning of the lan
guage model or reliance on thresholding tech
niques not available in practical settings1
1 Introduction
The advancement of cuttingedge Large Language
Models LLMs like GPT4 Claude 3 LLaMA2
and Gemini variants introduces a significant chal
lenge despite producing content that is linguis
tically coherent their outputs frequently contain
misleading or false information often referred to
as hallucinations or factual inconsistencies Hal
lucinations in Large Language Models refer to in
stances where the model generates usually plau
sible but entirely fabricated information Factual
inconsistencies a specific type of hallucination
1Code and data are available on GitHuboccur when generated text contradicts the source
material or other wellestablished facts not explic
itly mentioned in the source
Traditional automatic evaluation methodologies
like ROUGE Lin 2004 METEOR Banerjee and
Lavie 2005 and BLEU Papineni et al 2002
have been instrumental in assessing Natural Lan
guage Generation tasks However numerous stud
ies demonstrate the lack of correlation between ini
tial automatic evaluation models and human judg
ment in tasks such as machine translation Callison
Burch et al 2006 Bhattacharyya et al 2007
image captioning Cui et al 2018 and notably
factuality Fu et al 2023 Mao et al 2023 In par
ticular these models struggle to capture semantic
equivalence when there are substantial discrepan
cies in length syntax and wording between two
texts Guo and V osoughi 2023 Stent et al 2005
Consequently specialized models Laban et al
2021 Kry scinski et al 2019 Goyal and Durrett
2021 have been developed to assess textual factual
consistency verifying the truthfulness of a claim
or summary based on given ground truth textual
content
However existing models often finetuned vari
ants of RoBERTa Liu et al 2019 for assessing
factual consistency exhibit significant limitations
As highlighted in Tang et al 2023 these mod
els show reduced effectiveness in detecting fac
tual inconsistencies in content produced by recent
stateoftheart textgenerating models Ensemble
learning is the practice of merging the outputs of
multiple models to produce a more accurate predic
tion Dietterich 2000 Forbes et al 2023 demon
strated that ensembling factual consistency mod
els by calculating their weighted mean surpassed
the performance of individual models in detect
ing hallucinations within a small dataset of GPT3
generated Wikipedia abstractive summaries
In light of these limitations this study evaluates
benchmarks exclusively featuring summaries fromarXiv240613009v1 csCL 18 Jun 2024recent transformerbased language models This ap
proach more accurately reflects actual usage scenar
ios where users commonly need to validate texts
generated by newer LLMs rather than texts from
older text generation models We assess factual
consistency using the AggreFactXSUM FTSOTA
TofuEval SummaryLevel and HaluEval Summa
rization Tang et al 2023 2024 Li et al 2023
benchmarks consisting of transformergenerated
abstractive summaries featuring hallucinations that
existing models struggle to identify
Existing factual consistency encoder models out
put numerical scores requiring thresholding to map
the scores to binary labels However Tang et al
2023 demonstrates that the optimal threshold for
factual consistency models varies depending on the
recentness of the summarization model within the
AggreFact dataset Previous studies that report the
performance of these factual consistency models
have finetuned each models threshold using the
development subset of the same dataset under eval
uation Laban et al 2021 Fabbri et al 2022 Tang
et al 2023 2024
This approach is problematic and unrealistic as
it assumes access to labeled data from the target
dataset which may not be available in realworld
scenarios In this study we benchmark five popu