tensorkelechi/arxiv-cstext · Datasets at Hugging Face

text stringlengths 0 89.3k
consistent if all summarization sentences are indi
vidually marked as consistent5We thus refer to
this concatenation interchangeably as TofuEval
SummaryLevel or simply TofuEval Summary
We use the provided test subsets of each dataset
except for HaluEval Summarization which lacks a
traintest split For HaluEval Summarization we
use a balanced random sample of 3000 summaries
as our test set
3 Methodology
31 Ensembling Methods
We train and evaluate 16 ensembling methods
When applicable parameters for all relevant meth
ods are determined using a grid search over the
5Tang et al 2024 separated summaries into Main and
Marginal with the majority of sentences categorized as Main
We focus on the Main summaries because the Marginal dataset
in TofuEval is insufficiently small for reliable analysisparameter feature space Below is a listing of these
methods
Linear Models LogisticRegression LDA
TreeBased Methods RandomForest Gra
dientBoosting AdaBoost DecisionTree Cat
Boost XGB LGBM
Ensemble Voting MajorityLabelV oter
WeightedMajorityLabelV oter
Label Aggregation Models LabelModel
DawidSkene6
Other Methods Support Vector Machines
Nearest Neighbors Naive Bayes Bern
boulliNB
We consider MajorityLabelV oter as an ensem
bling baseline conceptually similar to averaging
binary scores with a threshold of 05 By com
paring the results of the majority vote to other en
sembling methods we can empirically assess the
additional performance gains achievable through
more sophisticated ensembling techniques
6The LabelModel as delineated in Ratner et al 2017
is particularly effective in learning the conditional probabili
ties of labeling functions adeptly reweighting their outputs in
semisupervised contexts The LabelModel represents an evo
lution in semisupervised learning encompassing techniques
such as those in FlyingSquid Fu et al 2020 DawidSkene
Dawid and Skene 2018 Data Programming Ratner et al
2016 and MeTaL Ratner et al 2019 all of which were
originally included as part of the Wrench benchmark Zhang
et al 202132 Metrics Used
Following Tang et al 2023 and Laban et al
2021 we use balanced accuracy to assess each
models proficiency in detecting factual errors Bal
anced accuracy helps provide a less biased evalua
tion by balancing the importance of sensitivity and
specificity ensuring that the predominance of the
majority class does not skew the results
We measure the reliability of predicted proba
bilities for factual consistency using Expected Cal
ibration Error ECE Expected Calibration Error
ECE is calculated by partitioning predictions into
nbins based on their confidence levels computing
the absolute difference between the actual accuracy
and the predicted probability in each bin and then
taking the weighted average of these differences
across all bins We use a ECE effectively measures
the discrepancy between a models confidence in its
predictions and its actual performance with lower
values indicating better calibration
ECE MX
m1Bm
NaccBmconf Bm1
Where Mis the number of bins Nis the total
number of samples Bmis the set of samples in
binmBmis the number of samples in bin m
accBmis the accuracy of predictions in bin m
andconf Bmis the average predicted probability
in bin m We selected M 8 for all ECE eval
uations as it offers reasonable balance between
statistical reliability and resolution
33 Calibrating Ensembled Models for
Reliable Probability Estimates
Histogram Binning Zadrozny and Elkan 2002
Bayesian Binning into Quantiles BBQ Naeini
et al 2015 and Isotonic Regression Zadrozny
and Elkan 2002 are nonparametric methods meth
ods of calibration while Temperature Scaling Guo
et al 2017 and Platt Scaling Platt 1999 are para
metric Platt Scaling applies a sigmoid function to
model outputs to calibrate them We test and ap
ply Platt Scaling BBQ Histogram Binning and
Isotonic Regression to obtain reliable probability
estimates from our ensembled models
34 Methodology for Prompt Creation
Our LLM prompts created using GPT4 in the
OpenAI playground employs various Chain of
Thought CoT Wei et al 2023 approaches toguide models through a structured evaluation of
factual consistency Most prompts have various
explicit evaluation criteria requiring the LLM to
determine if each claim in the summary can be
inferred directly from the context
35 Selection of Prompt Pool
Prompt selection for each prompt pool size was

consistent if all summarization sentences are indi

vidually marked as consistent5We thus refer to

this concatenation interchangeably as TofuEval

SummaryLevel or simply TofuEval Summary

We use the provided test subsets of each dataset

except for HaluEval Summarization which lacks a

traintest split For HaluEval Summarization we

use a balanced random sample of 3000 summaries

as our test set

3 Methodology

31 Ensembling Methods

We train and evaluate 16 ensembling methods

When applicable parameters for all relevant meth

ods are determined using a grid search over the

5Tang et al 2024 separated summaries into Main and

Marginal with the majority of sentences categorized as Main

We focus on the Main summaries because the Marginal dataset

in TofuEval is insufficiently small for reliable analysisparameter feature space Below is a listing of these

methods

Linear Models LogisticRegression LDA

TreeBased Methods RandomForest Gra

dientBoosting AdaBoost DecisionTree Cat

Boost XGB LGBM

Ensemble Voting MajorityLabelV oter

WeightedMajorityLabelV oter

Label Aggregation Models LabelModel

DawidSkene6

Other Methods Support Vector Machines

Nearest Neighbors Naive Bayes Bern

boulliNB

We consider MajorityLabelV oter as an ensem

bling baseline conceptually similar to averaging

binary scores with a threshold of 05 By com

paring the results of the majority vote to other en

sembling methods we can empirically assess the

additional performance gains achievable through

more sophisticated ensembling techniques

6The LabelModel as delineated in Ratner et al 2017

is particularly effective in learning the conditional probabili

ties of labeling functions adeptly reweighting their outputs in

semisupervised contexts The LabelModel represents an evo

lution in semisupervised learning encompassing techniques

such as those in FlyingSquid Fu et al 2020 DawidSkene

Dawid and Skene 2018 Data Programming Ratner et al

2016 and MeTaL Ratner et al 2019 all of which were

originally included as part of the Wrench benchmark Zhang

et al 202132 Metrics Used

Following Tang et al 2023 and Laban et al

2021 we use balanced accuracy to assess each

models proficiency in detecting factual errors Bal

anced accuracy helps provide a less biased evalua

tion by balancing the importance of sensitivity and

specificity ensuring that the predominance of the

majority class does not skew the results

We measure the reliability of predicted proba

bilities for factual consistency using Expected Cal

ibration Error ECE Expected Calibration Error

ECE is calculated by partitioning predictions into

nbins based on their confidence levels computing

the absolute difference between the actual accuracy

and the predicted probability in each bin and then

taking the weighted average of these differences

across all bins We use a ECE effectively measures

the discrepancy between a models confidence in its

predictions and its actual performance with lower

values indicating better calibration

ECE MX

m1Bm

NaccBmconf Bm1

Where Mis the number of bins Nis the total

number of samples Bmis the set of samples in

binmBmis the number of samples in bin m

accBmis the accuracy of predictions in bin m

andconf Bmis the average predicted probability

in bin m We selected M 8 for all ECE eval

uations as it offers reasonable balance between

statistical reliability and resolution

33 Calibrating Ensembled Models for

Reliable Probability Estimates

Histogram Binning Zadrozny and Elkan 2002

Bayesian Binning into Quantiles BBQ Naeini

et al 2015 and Isotonic Regression Zadrozny

and Elkan 2002 are nonparametric methods meth

ods of calibration while Temperature Scaling Guo

et al 2017 and Platt Scaling Platt 1999 are para

metric Platt Scaling applies a sigmoid function to

model outputs to calibrate them We test and ap

ply Platt Scaling BBQ Histogram Binning and

Isotonic Regression to obtain reliable probability

estimates from our ensembled models

34 Methodology for Prompt Creation

Our LLM prompts created using GPT4 in the

OpenAI playground employs various Chain of

Thought CoT Wei et al 2023 approaches toguide models through a structured evaluation of

factual consistency Most prompts have various

explicit evaluation criteria requiring the LLM to

determine if each claim in the summary can be

inferred directly from the context

35 Selection of Prompt Pool

Prompt selection for each prompt pool size was