tensorkelechi/arxiv-cstext · Datasets at Hugging Face

text stringlengths 0 89.3k
sistent with the provided knowledge and
dialogue history If there is a logical con
flict respond with inconsistent If there is
no conflict respond with consistent
Input Text
Dialogue History Please summarize
the given knowledge
Response
Luo et al 2023 ZeroShot
Decide if the following summary is consis
tent with the corresponding article Note
that consistency means all information in
the summary is supported by the article
Article
Summary
Answer yes or no
Luo et al 2023 ChainofThought
Decide if the following summary is consis
tent with the corresponding article Note
that consistency means all information in
the summary is supported by the article
Article
Summary
Explain your reasoning step by step then
answer yes or no the questionWang et al 2023 ChatGPTDA
Score the following news summarization
given the corresponding news with respect
to consistency on a continuous scale from 0
to 100 where a score of zero means incon
sistency and score of one hundred means
perfect consistency Note that consistency
measures whether the facts in the summary
are consistent with the facts in the original
article Consider whether the summary does
reproduce all facts accurately and does not
make up untrue information
Article
Summary
Scores
Wang et al 2023 ChatGPTStar
Score the following news summarization
given the corresponding news with respect
to consistency with one to five stars where
one star means inconsistency and five
stars means perfect consistency Note
that consistency measures whether the facts
in the summary are consistent with the facts
in the original article Consider whether the
summary does reproduce all facts accurately
and does not make up untrue information
Article
Summary
Stars
C Supplemental Results
Table 5 provide the full calibration results before
and after calibration across all datasets Table 6
shows the performance gained through ensembling
compared to using the best performing individual
promptModelAggreFactXSUM FTSOTA AggreFactCNNDM FTSOTA TofuEval MediaSum TofuEval MeetingBank
Uncal Platt BBQ Hist Isotonic Uncal Platt BBQ Hist Isotonic Uncal Platt BBQ Hist Isotonic Uncal Platt BBQ Hist Isotonic
AdaBoost 51 21 35 37 37 112 69 86 74 57 126 82 63 110 109 57 67 77 95 87
BernoulliNB 92 51 48 48 42 118 72 106 81 61 73 103 71 78 83 148 46 19 17 29
CatBoost 82 25 42 58 52 61 74 91 91 91 98 128 128 73 85 68 67 53 98 54
DecisionTree 89 64 66 66 46 39 29 49 75 39 58 108 80 89 58 52 99 82 77 52
GradientBoosting 75 46 58 71 41 63 93 98 93 73 98 112 89 91 87 53 94 26 98 63
KNeighbors 110 49 132 75 45 54 84 76 63 74 69 94 145 90 78 70 54 80 57 48
LabelModel 147 41 43 84 62 159 85 99 56 73 118 149 163 97 137 188 19 62 29 54
LDA 69 45 45 36 34 68 61 98 78 64 85 118 63 110 108 76 55 45 80 43
LGBM 155 49 60 58 53 110 60 73 62 55 131 103 113 107 95 73 49 32 65 56
LogisticRegression 134 49 57 57 57 57 67 96 85 63 108 111 61 109 98 43 54 73 73 73
MultinomialNB 83 81 78 79 79 129 68 102 90 67 101 47 50 71 67 177 54 49 32 40
RandomForest 98 63 70 99 60 109 126 154 105 100 132 197 173 128 130 84 163 63 147 92
SVC 93 09 111 57 83 58 107 169 111 78 134 148 117 135 110 54 105 111 81 53
XGB 72 64 88 124 65 119 69 230 167 124 145 82 144 224 129 82 49 69 81 69
Table 5 ECE comparison for ensembled models uncalibrated vs calibrated with Platt Scaling BBQ Histogram
Binning and Isotonic across datasets Highlighted text denotes the ensemblecalibration pair with the lowest ECE
per dataset
Dataset Best Individual Prompt Majority Label VoterBest Ensemble Model
Model Bal Acc
AggreFactXSUM FTSOTA 691 711 20 LabelModel 719 28
HaluEval Summarization 727 733 06 KNeighbors 749 22
TofuEvalSummaryLevelMediaSum 666 643 23 LabelModel 663 03
TofuEvalSummaryLevelMeetingBank 732 741 09 LabelModel 797 65
Table 6 Showcasing the improvement in balanced accuracy achieved by ensembling compared to the topperforming
individual prompts for each dataset The number in parentheses shows the difference in balanced accuracy between
the best individual prompt and the ensemble for each dataset MajorityLabelV oter serves as our baseline ensemble
offering a simple trainingfree method to combine results Numbers in all columns are sourced from Figure 2
for ensembling results and Figure 1 for individual prompt performance and are chosen by performance The
performance improvement in parentheses in the final column is the difference between the best ensemble model and
the individual prompt performance This represents an optimistic view of the possible the performance gain from
ensembling for each datasetHierarchical Compression of TextRich Graphs via Large
Language Models
Shichang Zhang
University of California Los Angeles
shichangcsuclaeduDa Zheng
Amazon
dzzhenamazoncomJiani Zhang
Amazon
zhajianiamazoncom
Qi Zhu

sistent with the provided knowledge and

dialogue history If there is a logical con

flict respond with inconsistent If there is

no conflict respond with consistent

Input Text

Dialogue History Please summarize

the given knowledge

Response

Luo et al 2023 ZeroShot

Decide if the following summary is consis

tent with the corresponding article Note

that consistency means all information in

the summary is supported by the article

Article

Summary

Answer yes or no

Luo et al 2023 ChainofThought

Decide if the following summary is consis

tent with the corresponding article Note

that consistency means all information in

the summary is supported by the article

Article

Summary

Explain your reasoning step by step then

answer yes or no the questionWang et al 2023 ChatGPTDA

Score the following news summarization

given the corresponding news with respect

to consistency on a continuous scale from 0

to 100 where a score of zero means incon

sistency and score of one hundred means

perfect consistency Note that consistency

measures whether the facts in the summary

are consistent with the facts in the original

article Consider whether the summary does

reproduce all facts accurately and does not

make up untrue information

Article

Summary

Scores

Wang et al 2023 ChatGPTStar

Score the following news summarization

given the corresponding news with respect

to consistency with one to five stars where

one star means inconsistency and five

stars means perfect consistency Note

that consistency measures whether the facts

in the summary are consistent with the facts

in the original article Consider whether the

summary does reproduce all facts accurately

and does not make up untrue information

Article

Summary

Stars

C Supplemental Results

Table 5 provide the full calibration results before

and after calibration across all datasets Table 6

shows the performance gained through ensembling

compared to using the best performing individual

promptModelAggreFactXSUM FTSOTA AggreFactCNNDM FTSOTA TofuEval MediaSum TofuEval MeetingBank

Uncal Platt BBQ Hist Isotonic Uncal Platt BBQ Hist Isotonic Uncal Platt BBQ Hist Isotonic Uncal Platt BBQ Hist Isotonic

AdaBoost 51 21 35 37 37 112 69 86 74 57 126 82 63 110 109 57 67 77 95 87

BernoulliNB 92 51 48 48 42 118 72 106 81 61 73 103 71 78 83 148 46 19 17 29

CatBoost 82 25 42 58 52 61 74 91 91 91 98 128 128 73 85 68 67 53 98 54

DecisionTree 89 64 66 66 46 39 29 49 75 39 58 108 80 89 58 52 99 82 77 52

GradientBoosting 75 46 58 71 41 63 93 98 93 73 98 112 89 91 87 53 94 26 98 63

KNeighbors 110 49 132 75 45 54 84 76 63 74 69 94 145 90 78 70 54 80 57 48

LabelModel 147 41 43 84 62 159 85 99 56 73 118 149 163 97 137 188 19 62 29 54

LDA 69 45 45 36 34 68 61 98 78 64 85 118 63 110 108 76 55 45 80 43

LGBM 155 49 60 58 53 110 60 73 62 55 131 103 113 107 95 73 49 32 65 56

LogisticRegression 134 49 57 57 57 57 67 96 85 63 108 111 61 109 98 43 54 73 73 73

MultinomialNB 83 81 78 79 79 129 68 102 90 67 101 47 50 71 67 177 54 49 32 40

RandomForest 98 63 70 99 60 109 126 154 105 100 132 197 173 128 130 84 163 63 147 92

SVC 93 09 111 57 83 58 107 169 111 78 134 148 117 135 110 54 105 111 81 53

XGB 72 64 88 124 65 119 69 230 167 124 145 82 144 224 129 82 49 69 81 69

Table 5 ECE comparison for ensembled models uncalibrated vs calibrated with Platt Scaling BBQ Histogram

Binning and Isotonic across datasets Highlighted text denotes the ensemblecalibration pair with the lowest ECE

per dataset

Dataset Best Individual Prompt Majority Label VoterBest Ensemble Model

Model Bal Acc

AggreFactXSUM FTSOTA 691 711 20 LabelModel 719 28

HaluEval Summarization 727 733 06 KNeighbors 749 22

TofuEvalSummaryLevelMediaSum 666 643 23 LabelModel 663 03

TofuEvalSummaryLevelMeetingBank 732 741 09 LabelModel 797 65

Table 6 Showcasing the improvement in balanced accuracy achieved by ensembling compared to the topperforming

individual prompts for each dataset The number in parentheses shows the difference in balanced accuracy between

the best individual prompt and the ensemble for each dataset MajorityLabelV oter serves as our baseline ensemble

offering a simple trainingfree method to combine results Numbers in all columns are sourced from Figure 2

for ensembling results and Figure 1 for individual prompt performance and are chosen by performance The

performance improvement in parentheses in the final column is the difference between the best ensemble model and

the individual prompt performance This represents an optimistic view of the possible the performance gain from

ensembling for each datasetHierarchical Compression of TextRich Graphs via Large

Language Models

Shichang Zhang

University of California Los Angeles

shichangcsuclaeduDa Zheng

Amazon

dzzhenamazoncomJiani Zhang

Amazon

zhajianiamazoncom

Qi Zhu