tensorkelechi/arxiv-cstext · Datasets at Hugging Face

text stringlengths 0 89.3k
lar stateoftheart factual consistency models and
demonstrate a substantial decline in performance
when thresholds are learned from different datasets
or set to their default midpoint
We extend the findings of Tang et al 2023 and
discover that the optimal threshold for each fac
tual consistency model varies widely across differ
ent datasets even when evaluating text generated
solely from recent summarization models Fig
ure 1 reveals that even when considering only
datasets with summaries from recent transformer
models the optimal linear threshold for maximiz
ing balanced accuracy differs widely for each fac
tual consistency model covering a broad range of
their possible output scores2Moreover Figure 2
demonstrates that optimizing thresholds on non
test data or setting them to the midpoint of each
models score range substantially reduces balanced
accuracy compared to testset optimization The
reliance of these models on datasetspecific thresh
olds as demonstrated by our findings limits their
practical utility in evaluating factual consistency
2QuestEval SummaCConv and AlignScore with mode
nli_sp generate scores from 0 to 1 SummaCZS scores
range from 1 to 1 QAFactEval scores range from 0 to 5across a diverse range of text without further fine
tuning or adjustments
Figure 1 Optimal thresholds of factual consistency
models when set to maximize balanced accuracy on
each test dataset
Previous efforts to use LLMs for identifying fac
tual inconsistencies include Wang et al 2023
who prompted ChatGPT to return numerical fac
tuality scores for summarization and Luo et al
2023 who used ChatGPT to produce binary fac
tuality judgments However Tang et al 2023
showed that these existing prompts had shown poor
performance on detecting factual consistencies over
the AggreFact FTSOTA benchmark of transformer
generated summaries However reliably mapping
the generated token probabilities from the reason
ing steps to the final answers confidence remains
unclear especially considering factors like temper
ature and sampling methods that influence token
generation
Existing LLM solutions are frequently overcon
fident in their assessments of a texts factual con
sistency Tang et al 2024 showed that their
summarylevel binary factual consistency prompts
when used with GPT4 frequently failed to identify
sentences with factual errors resulting in False Pos
itive Rates of 69 and 46 on the MediaSum and
MeetingBank subsets of the TofuEval Summary
Level dataset respectively This issue of overcon
fidence is not unique to their approach but rather
a general problem that neural networks including
LLMs tend to be overconfident in their predictions
Guo et al 2017 Minderer et al 2021 Jiang et al
2021 Xiong et al 20233
3Binary prompts by their very nature force models to
reduce nuance and uncertainty into a single decision conFigure 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 mid
point of each models score range Optimizing at Cen
ter or optimizing on all datasets except the test set
Optimizing on Train which reflects a realistic sce
nario of applying the model to unseen data Numbers
above bars quantify the decrease in balanced accuracy
when thresholds are optimized on nontest data com
pared to the test dataset itself underscoring the difficulty
of effectively applying these models to unseen data in
practice
Confidence elicitation an increasingly popular
method involves prompting the LLM to output
its uncertainty along with its prediction Lin et al
2022 Xiong et al 2023 Tian et al 2023 Despite
its potential no confidence elicitation approach has
yet consistently yielded accurate confidence esti
mates across diverse LLMs and tasks limiting its
current practical utility Calibration the process of
adjusting modelpredicted probabilities to match
their empirical accuracies has remained the lead
ing solution to correct model overconfidence Guo
et al 2017
We propose a novel approach crafting a di
verse set of LLM prompts that each output a binary
score indicating each prompts belief of whether
the summaries contain any factual errors These bi
tributing to the overconfidence observed in their responsesnary features are then fed into ensembling models
which integrate the multiple perspectives of each
prompt to produce a single probability Finally we
calibrate the ensemble models to obtain empirically
accurate probabilities that a given summary is fac
tually consistent or free of hallucination4The full
pipeline of our framework can be seen in Figure 3
The primary contributions of our work can be
outlined as follows 1We demonstrate that prior
methods for detecting factual errors in summariza

lar stateoftheart factual consistency models and

demonstrate a substantial decline in performance

when thresholds are learned from different datasets

or set to their default midpoint

We extend the findings of Tang et al 2023 and

discover that the optimal threshold for each fac

tual consistency model varies widely across differ

ent datasets even when evaluating text generated

solely from recent summarization models Fig

ure 1 reveals that even when considering only

datasets with summaries from recent transformer

models the optimal linear threshold for maximiz

ing balanced accuracy differs widely for each fac

tual consistency model covering a broad range of

their possible output scores2Moreover Figure 2

demonstrates that optimizing thresholds on non

test data or setting them to the midpoint of each

models score range substantially reduces balanced

accuracy compared to testset optimization The

reliance of these models on datasetspecific thresh

olds as demonstrated by our findings limits their

practical utility in evaluating factual consistency

2QuestEval SummaCConv and AlignScore with mode

nli_sp generate scores from 0 to 1 SummaCZS scores

range from 1 to 1 QAFactEval scores range from 0 to 5across a diverse range of text without further fine

tuning or adjustments

Figure 1 Optimal thresholds of factual consistency

models when set to maximize balanced accuracy on

each test dataset

Previous efforts to use LLMs for identifying fac

tual inconsistencies include Wang et al 2023

who prompted ChatGPT to return numerical fac

tuality scores for summarization and Luo et al

2023 who used ChatGPT to produce binary fac

tuality judgments However Tang et al 2023

showed that these existing prompts had shown poor

performance on detecting factual consistencies over

the AggreFact FTSOTA benchmark of transformer

generated summaries However reliably mapping

the generated token probabilities from the reason

ing steps to the final answers confidence remains

unclear especially considering factors like temper

ature and sampling methods that influence token

generation

Existing LLM solutions are frequently overcon

fident in their assessments of a texts factual con

sistency Tang et al 2024 showed that their

summarylevel binary factual consistency prompts

when used with GPT4 frequently failed to identify

sentences with factual errors resulting in False Pos

itive Rates of 69 and 46 on the MediaSum and

MeetingBank subsets of the TofuEval Summary

Level dataset respectively This issue of overcon

fidence is not unique to their approach but rather

a general problem that neural networks including

LLMs tend to be overconfident in their predictions

Guo et al 2017 Minderer et al 2021 Jiang et al

2021 Xiong et al 20233

3Binary prompts by their very nature force models to

reduce nuance and uncertainty into a single decision conFigure 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 mid

point of each models score range Optimizing at Cen

ter or optimizing on all datasets except the test set

Optimizing on Train which reflects a realistic sce

nario of applying the model to unseen data Numbers

above bars quantify the decrease in balanced accuracy

when thresholds are optimized on nontest data com

pared to the test dataset itself underscoring the difficulty

of effectively applying these models to unseen data in

practice

Confidence elicitation an increasingly popular

method involves prompting the LLM to output

its uncertainty along with its prediction Lin et al

2022 Xiong et al 2023 Tian et al 2023 Despite

its potential no confidence elicitation approach has

yet consistently yielded accurate confidence esti

mates across diverse LLMs and tasks limiting its

current practical utility Calibration the process of

adjusting modelpredicted probabilities to match

their empirical accuracies has remained the lead

ing solution to correct model overconfidence Guo

et al 2017

We propose a novel approach crafting a di

verse set of LLM prompts that each output a binary

score indicating each prompts belief of whether

the summaries contain any factual errors These bi

tributing to the overconfidence observed in their responsesnary features are then fed into ensembling models

which integrate the multiple perspectives of each

prompt to produce a single probability Finally we

calibrate the ensemble models to obtain empirically

accurate probabilities that a given summary is fac

tually consistent or free of hallucination4The full

pipeline of our framework can be seen in Figure 3

The primary contributions of our work can be

outlined as follows 1We demonstrate that prior

methods for detecting factual errors in summariza