tensorkelechi/arxiv-cstext · Datasets at Hugging Face

text stringlengths 0 89.3k
generally outperforms the other tested calibration
methods achieving postcalibration ECE scores of
under 7
Reliability diagrams are tools for assessing the
accuracy of probability estimates from both uncali
brated and calibrated ensemble models The relia
bility diagram in Figure 4 shows that Platt Scaling
when applied to an ensemble model can signifi
cantly reduce overconfidence in predicting a texts
factual consistency
44 Statistical Testing
We assess if the performance gains in LLM en
sembling over existing methods in models are sta
tistically significant For the AggreFactXSUM
FTSOTA TofuEval and HaluEval Summarization
dataset we execute distinct evaluations using boot
strap resampling techniques Efron 1982 com
paring the best Ensembling Method against the
previous top performing model for each dataset
Following Laban et al 2021 we conduct compar
isons at a statistical significance level of p 001
incorporating the Bonferroni adjustment Bonfer
roni 1935 due to the multiple tests conducted
on the datasets Our LLM ensembling methods
demonstrate statistically meaningful advancements
on the HaluEval dataset with statistically signif
icant pvalues of under 01 in both the original
and Bonferroniadjusted analyses However the
statistical tests do not confirm a statistically signifi
cant advantage with pvalues below 01 comparedEnsemblingAggreFact HaluEval TofuEval MediaSum TofuEval MeetingBank
MethodXSUM FTSOTA Summarization SummaryLevel SummaryLevel
Number of Prompts Number of Prompts Number of Prompts Number of Prompts
3 5 9 3 5 9 3 5 9 3 5 9
Baseline 5000 5000 5000 5000 5000 5000 5000 5000 5000 5000 5000 5000
AdaBoost 6704 6704 7021 7147 7373 7267 6580 6380 6474 7117 7289 7197
BernoulliNB 6704 7102 6807 7327 7233 6857 6424 6596 6480 7407 7638 7716
CatBoost 6704 6704 6873 7147 7180 7153 6580 6380 6474 7117 7289 7380
DawidSkene 6917 6962 7028 7327 7053 6737 6417 6433 6347 7407 7370 7191
DecisionTree 6704 6704 6611 7327 6867 7150 6016 6490 6331 7316 7391 6700
GradientBoosting 6704 6653 6891 7390 7107 6763 6580 6467 6430 7117 7391 7380
KNeighbors 6704 6704 6739 7487 7143 6877 6424 6417 6328 7407 6955 7108
LDA 6704 6704 6979 7327 7383 7060 6580 6380 6444 7117 7289 7103
LabelModel 6940 7192 6793 7390 7167 6850 6424 6633 6553 7410 7938 7974
LGBM 6704 6704 6873 7390 7407 7343 6580 6467 6517 7407 7391 7358
LogisticRegression 6704 6704 7017 7327 7253 7040 6580 6380 6377 7117 7289 7074
MajorityLabelV oter 6917 6962 7105 7327 7037 6787 6417 6433 6350 7410 7370 7067
RandomForest 6704 6672 6791 7390 7087 7077 6580 6490 6354 7117 7399 7358
SVC 6704 6672 6761 7390 7287 7203 6580 6467 6430 7117 7391 7227
WeightedMajorityV oting 6704 6704 7017 7327 7373 7260 6580 6380 6377 7117 7289 7074
XGB 6704 6672 7039 7390 7173 6673 6580 6490 6387 7117 7399 7214
Table 2 Exploring the impact of various LLM prompt sizes and ensembling methods on balanced accuracy across
all test datasets The top performing ensemble method for each prompt sizedataset combination is shown in bold
Ensembling as few as three prompts consistently yields performance improvements across all datasets compared to
the bestperforming individual prompt
Method Type Method AggreFact HaluEval TofuEval MediaSum TofuEval MeetingBank
XSUM FTSOTA Summarization SummaryLevel SummaryLevel
Encoder ModelsAlignScore 70238 65807 62358 70154
QuestEval 51342 55422 53760 50859
SummaCZS 53042 60507 48660 57459
SummaCCv 50942 58707 53359 45261
QAFactEval 62440 52022 54659 57857
LLM SolutionsChatGPTZS GPT35 62140 64318 63658 69053
ChatGPTCOT GPT35 55441 62518 63158 66852
ChatGPTDA GPT35 56441 59618 52759 52358
ChatGPTStar GPT35 55441 61518 57458 55757
Tang2024Summary GPT35 62440 64009 61934 71931
Tang2024Summary GPT4 62940 66109 62334 72931
LLM EnsemblesEnsembleTop3 GPT4 69438 74916 65857 74151
EnsembledTop5 GPT4 71938 74116 66357 79451
EnsembleTop9 Mixed 71138 73416 65557 79749
Table 3 A chart comparing the balanced accuracy of encoderbased models and LLM solutions in identifying
factual inconsistencies and hallucinations For each test dataset the encoderbased model scores are obtained using
linear thresholds optimized on the other three datasets ensuring that neither the test data nor its validation set is
used for threshold tuning Lower performance in encoder models compared to existing studies is due to evaluating
without finetuning each models threshold using the test datasets development subset 95 confidence intervals
are shown with the highest performing method for each dataset in bold Mixed refers to ensembling the binary
LLM outputs from both GPT35 and GPT4 The existing prompts and relevant citations use in Method Type LLM
Solutions can be found in Appendix Section B
to the secondbest methods across the remaining
three datasets The limited number of samples
in the AggreFactXSUM FTSOTA and TofuEval
datasets in contrast to the larger HaluEval dataset
necessitates a significantly greater improvement in
performance to achieve statistical significance11
11To eliminate any biases caused by varying sample sizes
across models we exclusively conduct statistical testing on
the overlapping subsets of all samples under consideration5 Conclusion
We introduce Detecting Errors through Ensembling
Prompts DEEP a stateoftheart LLMbased
method for detecting factual consistencies and hal
lucinations in summaries Our findings reveal that
factual consistency encoder models exhibit a pro
thereby ensuring that differing sample sizes do not affect our
statistical significance testingModelAggreFactXSUM FTSOTA
3 Prompts 5 Prompts 9 Prompts
Uncal Platt Uncal Platt Uncal Platt
AdaBoost 51 21 120 57 187 43
BernoulliNB 92 51 159 50 215 72
CatBoost 82 25 81 52 77 67
DecisionTree 89 64 72 53 93 49

generally outperforms the other tested calibration

methods achieving postcalibration ECE scores of

under 7

Reliability diagrams are tools for assessing the

accuracy of probability estimates from both uncali

brated and calibrated ensemble models The relia

bility diagram in Figure 4 shows that Platt Scaling

when applied to an ensemble model can signifi

cantly reduce overconfidence in predicting a texts

factual consistency

44 Statistical Testing

We assess if the performance gains in LLM en

sembling over existing methods in models are sta

tistically significant For the AggreFactXSUM

FTSOTA TofuEval and HaluEval Summarization

dataset we execute distinct evaluations using boot

strap resampling techniques Efron 1982 com

paring the best Ensembling Method against the

previous top performing model for each dataset

Following Laban et al 2021 we conduct compar

isons at a statistical significance level of p 001

incorporating the Bonferroni adjustment Bonfer

roni 1935 due to the multiple tests conducted

on the datasets Our LLM ensembling methods

demonstrate statistically meaningful advancements

on the HaluEval dataset with statistically signif

icant pvalues of under 01 in both the original

and Bonferroniadjusted analyses However the

statistical tests do not confirm a statistically signifi

cant advantage with pvalues below 01 comparedEnsemblingAggreFact HaluEval TofuEval MediaSum TofuEval MeetingBank

MethodXSUM FTSOTA Summarization SummaryLevel SummaryLevel

Number of Prompts Number of Prompts Number of Prompts Number of Prompts

3 5 9 3 5 9 3 5 9 3 5 9

Baseline 5000 5000 5000 5000 5000 5000 5000 5000 5000 5000 5000 5000

AdaBoost 6704 6704 7021 7147 7373 7267 6580 6380 6474 7117 7289 7197

BernoulliNB 6704 7102 6807 7327 7233 6857 6424 6596 6480 7407 7638 7716

CatBoost 6704 6704 6873 7147 7180 7153 6580 6380 6474 7117 7289 7380

DawidSkene 6917 6962 7028 7327 7053 6737 6417 6433 6347 7407 7370 7191

DecisionTree 6704 6704 6611 7327 6867 7150 6016 6490 6331 7316 7391 6700

GradientBoosting 6704 6653 6891 7390 7107 6763 6580 6467 6430 7117 7391 7380

KNeighbors 6704 6704 6739 7487 7143 6877 6424 6417 6328 7407 6955 7108

LDA 6704 6704 6979 7327 7383 7060 6580 6380 6444 7117 7289 7103

LabelModel 6940 7192 6793 7390 7167 6850 6424 6633 6553 7410 7938 7974

LGBM 6704 6704 6873 7390 7407 7343 6580 6467 6517 7407 7391 7358

LogisticRegression 6704 6704 7017 7327 7253 7040 6580 6380 6377 7117 7289 7074

MajorityLabelV oter 6917 6962 7105 7327 7037 6787 6417 6433 6350 7410 7370 7067

RandomForest 6704 6672 6791 7390 7087 7077 6580 6490 6354 7117 7399 7358

SVC 6704 6672 6761 7390 7287 7203 6580 6467 6430 7117 7391 7227

WeightedMajorityV oting 6704 6704 7017 7327 7373 7260 6580 6380 6377 7117 7289 7074

XGB 6704 6672 7039 7390 7173 6673 6580 6490 6387 7117 7399 7214

Table 2 Exploring the impact of various LLM prompt sizes and ensembling methods on balanced accuracy across

all test datasets The top performing ensemble method for each prompt sizedataset combination is shown in bold

Ensembling as few as three prompts consistently yields performance improvements across all datasets compared to

the bestperforming individual prompt

Method Type Method AggreFact HaluEval TofuEval MediaSum TofuEval MeetingBank

XSUM FTSOTA Summarization SummaryLevel SummaryLevel

Encoder ModelsAlignScore 70238 65807 62358 70154

QuestEval 51342 55422 53760 50859

SummaCZS 53042 60507 48660 57459

SummaCCv 50942 58707 53359 45261

QAFactEval 62440 52022 54659 57857

LLM SolutionsChatGPTZS GPT35 62140 64318 63658 69053

ChatGPTCOT GPT35 55441 62518 63158 66852

ChatGPTDA GPT35 56441 59618 52759 52358

ChatGPTStar GPT35 55441 61518 57458 55757

Tang2024Summary GPT35 62440 64009 61934 71931

Tang2024Summary GPT4 62940 66109 62334 72931

LLM EnsemblesEnsembleTop3 GPT4 69438 74916 65857 74151

EnsembledTop5 GPT4 71938 74116 66357 79451

EnsembleTop9 Mixed 71138 73416 65557 79749

Table 3 A chart comparing the balanced accuracy of encoderbased models and LLM solutions in identifying

factual inconsistencies and hallucinations For each test dataset the encoderbased model scores are obtained using

linear thresholds optimized on the other three datasets ensuring that neither the test data nor its validation set is

used for threshold tuning Lower performance in encoder models compared to existing studies is due to evaluating

without finetuning each models threshold using the test datasets development subset 95 confidence intervals

are shown with the highest performing method for each dataset in bold Mixed refers to ensembling the binary

LLM outputs from both GPT35 and GPT4 The existing prompts and relevant citations use in Method Type LLM

Solutions can be found in Appendix Section B

to the secondbest methods across the remaining

three datasets The limited number of samples

in the AggreFactXSUM FTSOTA and TofuEval

datasets in contrast to the larger HaluEval dataset

necessitates a significantly greater improvement in

performance to achieve statistical significance11

11To eliminate any biases caused by varying sample sizes

across models we exclusively conduct statistical testing on

the overlapping subsets of all samples under consideration5 Conclusion

We introduce Detecting Errors through Ensembling

Prompts DEEP a stateoftheart LLMbased

method for detecting factual consistencies and hal

lucinations in summaries Our findings reveal that

factual consistency encoder models exhibit a pro

thereby ensuring that differing sample sizes do not affect our

statistical significance testingModelAggreFactXSUM FTSOTA

3 Prompts 5 Prompts 9 Prompts

Uncal Platt Uncal Platt Uncal Platt

AdaBoost 51 21 120 57 187 43

BernoulliNB 92 51 159 50 215 72

CatBoost 82 25 81 52 77 67

DecisionTree 89 64 72 53 93 49