tensorkelechi/arxiv-cstext · Datasets at Hugging Face

text stringlengths 0 89.3k
GradientBoosting 75 46 61 52 86 54
KNeighbors 110 49 104 47 98 45
LabelModel 147 41 238 47 228 53
LDA 69 45 74 61 71 41
LGBM 155 49 200 77 159 61
LogisticRegression 134 49 65 58 66 39
MultinomialNB 83 81 42 26 44 26
RandomForest 98 63 60 41 98 45
SVC 93 09 84 65 95 57
XGB 72 64 69 48 160 69
Table 4 Comparison of the Expected Calibration Error ECE for ensembling models before uncalibrated and
after calibration using Platt Scaling across various models and prompt pools The calibration model was trained
only on the three nontest datasets
06 08 10
Confidence0001020304050607080910Uncalibrated
P ositive P r edictions
ECE 2635
06 08 10
Confidence0001020304050607080910Uncalibrated
Negative P r edictions
ECE 1777
06 08 10
Confidence0001020304050607080910Calibrated
P ositive P r edictions
ECE 595
06 08 10
Confidence0001020304050607080910Calibrated
Negative P r edictions
ECE 430A ccuracyExample R eliability Diagram Applying Platt Scaling
to LabelModel with Nine P r ompts
Empirical A ccuracy Output Confidence
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 nontest datasets and tested on the AggreFactXSUM FTSOTA test dataset
nounced sensitivity to threshold settings with their
performance markedly declining if the threshold is
not adjusted based on the test datasets We demonstrate that DEEP surpasses the performance of ex
isting methods and models in evaluating the factual
consistency of summaries produced by recent transformer models Finally we show that by calibrating
ensemble models using binary input features de
rived from LLM prompts DEEP achieves reliable
probabilities indicating a texts factual consistency
or presence of errors
Limitations
LLM approaches for identifying factual errors are
significantly more resourceintensive than existing
finetuned encoder models requiring up to three
orders of magnitude more parameters Our method
requires more computation however this cost may
be justified in highstakes situations where failing
to identify factual errors is costly
Future research should explore the performance
of these prompts on more powerful language mod
els as they become available Additionally future
work should create a dataset of quality examples
of chainofthought reasoning to identify factual er
rors enabling fewshot learning to boost LLM per
formance Future work should consider comparing
the performance of ensembling factual consistency
model scores to LLM prompt ensembling
Future work should investigate why encoder
models for factual consistency evaluation require
datasetspecific linear thresholding for optimal per
formance Possible reasons include 1 the depen
dence of optimal binary classification thresholds on
the average number of sentences per summary as
most models calculate summary factuality scores
by averaging sentencelevel scores 2 dataset im
balances in the ratio of summaries with factual
errors with thresholding balancing specificity and
sensitivity and 3 inherent differences in iden
tifying factual errors across datasets even when
controlling for summary length and annotation im
balances due to factors like summarization models
or the contexts source However these reasons are
speculative and further experiments are needed to
test these hypotheses and explore approaches that
reduce model sensitivity to dataset characteristics
It is uncertain how many labeling errors exist in
the AggreFactXSUMSOTA and HaluEval Sum
marization datasets Future datasets should create
summarization error datasets where like TofuEval
ground truth labels are generated by combining
multiple humans annotations Additionally the
datasets under test contained solely English Sum
maries To evaluate LLMs ability to identify errors
across languages there is a need for multilingual
summarization error datasetsFinetuning LLMs for error detection and ap
plying our endtoend pipeline for spotting wider
ranges of language generation errors including
those in QA and machine translation should be
explored
References
Satanjeev Banerjee and Alon Lavie 2005 METEOR
An automatic metric for MT evaluation with im
proved correlation with human judgments In Pro
ceedings of the ACL Workshop on Intrinsic and Ex
trinsic Evaluation Measures for Machine Transla

GradientBoosting 75 46 61 52 86 54

KNeighbors 110 49 104 47 98 45

LabelModel 147 41 238 47 228 53

LDA 69 45 74 61 71 41

LGBM 155 49 200 77 159 61

LogisticRegression 134 49 65 58 66 39

MultinomialNB 83 81 42 26 44 26

RandomForest 98 63 60 41 98 45

SVC 93 09 84 65 95 57

XGB 72 64 69 48 160 69

Table 4 Comparison of the Expected Calibration Error ECE for ensembling models before uncalibrated and

after calibration using Platt Scaling across various models and prompt pools The calibration model was trained

only on the three nontest datasets

06 08 10

Confidence0001020304050607080910Uncalibrated

P ositive P r edictions

ECE 2635

06 08 10

Confidence0001020304050607080910Uncalibrated

Negative P r edictions

ECE 1777

06 08 10

Confidence0001020304050607080910Calibrated

P ositive P r edictions

ECE 595

06 08 10

Confidence0001020304050607080910Calibrated

Negative P r edictions

ECE 430A ccuracyExample R eliability Diagram Applying Platt Scaling

to LabelModel with Nine P r ompts

Empirical A ccuracy Output Confidence

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 nontest datasets and tested on the AggreFactXSUM FTSOTA test dataset

nounced sensitivity to threshold settings with their

performance markedly declining if the threshold is

not adjusted based on the test datasets We demonstrate that DEEP surpasses the performance of ex

isting methods and models in evaluating the factual

consistency of summaries produced by recent transformer models Finally we show that by calibrating

ensemble models using binary input features de

rived from LLM prompts DEEP achieves reliable

probabilities indicating a texts factual consistency

or presence of errors

Limitations

LLM approaches for identifying factual errors are

significantly more resourceintensive than existing

finetuned encoder models requiring up to three

orders of magnitude more parameters Our method

requires more computation however this cost may

be justified in highstakes situations where failing

to identify factual errors is costly

Future research should explore the performance

of these prompts on more powerful language mod

els as they become available Additionally future

work should create a dataset of quality examples

of chainofthought reasoning to identify factual er

rors enabling fewshot learning to boost LLM per

formance Future work should consider comparing

the performance of ensembling factual consistency

model scores to LLM prompt ensembling

Future work should investigate why encoder

models for factual consistency evaluation require

datasetspecific linear thresholding for optimal per

formance Possible reasons include 1 the depen

dence of optimal binary classification thresholds on

the average number of sentences per summary as

most models calculate summary factuality scores

by averaging sentencelevel scores 2 dataset im

balances in the ratio of summaries with factual

errors with thresholding balancing specificity and

sensitivity and 3 inherent differences in iden

tifying factual errors across datasets even when

controlling for summary length and annotation im

balances due to factors like summarization models

or the contexts source However these reasons are

speculative and further experiments are needed to

test these hypotheses and explore approaches that

reduce model sensitivity to dataset characteristics

It is uncertain how many labeling errors exist in

the AggreFactXSUMSOTA and HaluEval Sum

marization datasets Future datasets should create

summarization error datasets where like TofuEval

ground truth labels are generated by combining

multiple humans annotations Additionally the

datasets under test contained solely English Sum

maries To evaluate LLMs ability to identify errors

across languages there is a need for multilingual

summarization error datasetsFinetuning LLMs for error detection and ap

plying our endtoend pipeline for spotting wider

ranges of language generation errors including

those in QA and machine translation should be

explored

References

Satanjeev Banerjee and Alon Lavie 2005 METEOR

An automatic metric for MT evaluation with im

proved correlation with human judgments In Pro

ceedings of the ACL Workshop on Intrinsic and Ex

trinsic Evaluation Measures for Machine Transla