Datasets:
tensorkelechi
/
arxiv-cstext

Dataset card Data Studio Files Community
1
text
stringlengths
0
89.3k
tions perform significantly worse without the com
mon practice of an optimized threshold on sub
sets of the dataset under test 2We introduce a
large and diverse set of prompts each employing
unique methods and evaluation protocols to detect
hallucinations and factual inconsistencies in gen
erated summaries 3Our endtoend framework
achieves stateoftheart balanced accuracy on the
AggreFactXSUM FTSOTA TofuEval Summary
Level and HaluEval Summarization benchmarks
in detecting factual errors in transformergenerated
text summaries all without finetuning the lan
guage model or relying on impractical thresholding
techniques
2 Datasets
The AggreFact FTSOTA benchmark Tang et al
2023 tests a models ability to identify factual
inconsistencies in summaries produced by fine
tuned transformerbased summarization models
The dataset combines nine existing annotated fac
tuality datasets converting all factual consistency
scores to binary AggreFact is categorized based
on the development timeline of the underlying sum
marization models into FTSOTA EXFORMER
and OLD categories and is divided into AggreFact
CNNDM and AggreFactXSUM subsets We
benchmark AggreFactXSUM FTSOTA and ex
clude its CNNDM counterpart due to its insuf
ficient number of factual inconsistencies and its
more extractive summary style compared to the
more abstractive style of current SOTA LLMs
HaluEval Li et al 2023 a comprehensive
benchmark for assessing hallucination in LLMs
uses a ChatGPTbased twostep samplingthen
filtering framework to create the dataset We focus
4We craft prompts that aim to detect either factual incon
sistencies or hallucinations allowing the ensembling model
to determine the importance of each prompt for the datasets it
is trained on This approach enables us to sidestep the subtle
distinction between factual inconsistencies and hallucinations
and instead let the ensembling model decide what prompts are
most relevant for the training datan unique
prompts
Ensembler
Combines list of
boolean factuality
scores into one
probability 0
1
0
0
0LLM Prompt
Template 1
Determine if the
summary is
consistent or
inconsistent with the
provided context
Context
Summary
Output 0 if the text is
inconsistent and 1 if
the text is
consistent 1D Binary
Vector
Concatenating
Output of LLM
Prompts Calibrator
Adjusts predicted
probabilities to
match empirical
results
Summary
The Spurs beat the
Rockets 110 to 105 Context
In a close overtime
game the Rockets beat
the Spurs with a final
score of 110 to 105
Output
Probability that the
summary is factually
consistent or absent
of hallucination
Figure 3 Diagram of our endtoend framework
on the summarization subset of HaluEval which
pairs each of the 10000 sampled document con
texts with two summaries one with hallucination
and one without
TofuEval Tang et al 2024 is a topicfocused di
alogue summarization benchmark containing 15K
LLMgenerated summaries from the MediaSum
Zhu et al 2021 and MeetingBank Hu et al
2023 datasets TofuEval uses a twostage annota
tion process where two expert linguists indepen
dently assess the binary relevance completeness
and factual consistency of each sentence in the sum
maries We focus solely on the factual consistency
annotations of main topic summaries merging sum
marization sentences into one paragraph deemed