tensorkelechi/arxiv-cstext · Datasets at Hugging Face

text stringlengths 0 89.3k
While peers can leave a review only on submission length with out much eﬀort the score given by the
instructor usually does not depend on the cheap signal of len gth Thus the states θ θ1θm
are not equally important for scoring textual report qualit y and the less important ones need to
be ﬁltered
Our ﬁltering method identiﬁes important summary points for scoring by partitioning the states
by semantic meanings into topics t t1tz For example three states θ1θ2θ3for correct
hypothesis base case and induction step in proof can be gro uped into topic t1for the proof
correctness State θ4for typos can be grouped into topic t2of the overall clarity Intuitively
the number of diﬀerent indicator states within a topic is posi tively related to a high cost to obtain
information on that topic The ﬁltering approach ignores re ports on the topics with fewer summary
points instead scoring only the top 2 topics with the most su mmary points ie the most dimensions
to report
The ﬁltered average aggregation AF with top 2 topics is deﬁ ned as the following
Deﬁnition 9 Filtered Average AF Given scoring rules S1S z andi1andi2as the indices
of the remaining topics after ﬁltering A ﬁltered average ag gregation outputs S1
2Si1Si2
In the deﬁnition zis the total number of topics and S1S zare multidimensional scoring
rules that can be applied to multiple indicator states in eac h topic
9222 Textual Scoring Rules We Test
For textual scoring we use the Vshaped single dimensional scoring rules for each indicator state
on terminal nodes in the aggregation hierarchy
AV The average scoring rule of Vshaped on each indicator state
AMV The average of the maxoverseparate scoring rule applied t o each topic
AFV The average scoring rule of Vshaped ﬁltered to the top 2 dive rse topics
AFMV The average of the maxoverseparate scoring rule ﬁltered t o the top 2 diverse topics
3 ElicitationGPT Scoring Text with Oracle Access to LLM
In this section we present ElicitationGPT an implementation of scoring rules for text via queries
to large language models LLMs Instead of presenting deta ils of the prompts we model LLMs
as oracles with the fundamental functionalities of summarization andquestion answering We
present the algorithm via queries to these oracles This ora cle abstraction allows the main ideas
of the algorithm and its properties to be studied without det ails of prompt engineering tricks
in implementation For example the properness of the textu al elicitation mechanism is clearly
inherited from the proper scoring rule from which it is const ructed
In our oracle model the LLM is able to accept queries to two fu ndamental tasks In the
literature of natural language processing these tasks are also fundamental benchmarks to evaluate
LLM performance We refer to oracle queries as OSandOAas deﬁned below
Summarization OS
Input a list Iiiof documents
Output a list T1T2Tm of summary points
Our deﬁnition of the summarization task is a variant of the cl assic summarization task and
it is also known as the key point summarization task BarHai m et al 2020
Question Answering OA
Input text review Rand a list of text statements T1Tm
Output a vector r 01mof summarization points on T1Tm On each dimension
ri 1 ifRhas a segment supporting a positive opinion on Ti 0 if a negative opinion or if
Rdoes not mention anything related to Ti
The query for OAis DoesRsupport a positive opinion on the following statements T1Tm
Datasets for question answering include Rajpurkar et al 2 016 Kwiatkowski et al 2019
etc
Deﬁnition 10 ElicitationGPTis deﬁned from a large language model that gives summarizati on
and question answering oracles OSandOAand an knowitornot scoring rule S
Input
a list of ground truth text I1Iswithin the same cluster eg instructor reviews on
submissions to the same homework assignment
10an agent report Pcorresponding to the kth ground truth eg a peer review on the same
submission as the kth instructor review
Output A score in 01on the agent report
1extracting dimensions of summary points
Query summarization oracle and get a list of summary points f rom the reviews
T1Tm OSI1Is
2clustering topics optional
Cluster the summary points by similar topics Required for AF MV and AFV scoring rule
Topics t1tz OST1Tm
3calculating prior
Process the ground truths within the same cluster to determi ne the prior distribution over the
semantic state
For each ground truth text Ii determine the indicator state for each summary point
θiOAIiT1Tm
For each indicator state θj count the frequency pθjof1s from ground truth2
4mapping agent report to belief Construct the mapping from the agent report to beliefs on
the true semantic state
rOART1Tm
5knowitornot scoring rule Apply the scoring rule Sprθk
31 Properness of ElicitationGPT
An important property of numerical scoring rules is propern ess ie incentivizing the forecaster to
report their true beliefs In this section we give two result s about the incentives of ElicitationGPT
First we consider the perfect language oracle model and sh ow that in this model ElicitationGPTis
proper ie the agents optimal strategy is to report thei r true belief about how the instructor would
review the submission Second we show that if the agent does not exert eﬀort ie in peer grading
does not look at the submission then their score for any pos sible adversarial manipulation and
for imperfect language oracles is upper bounded by the scor e from reporting the prior mean
which they can easily obtain by reporting I dont know
Theorem 1 ElicitationGPTwith perfect language oracles is proper for knowitornot beliefs ie
the optimal strategy of an agent who knows the prior and has kn owitornot beliefs is to report an
accurate textual representation of their belief
2We mark 1 as a positive opinion on a summary point For each ind icator states and for ground truth text there
are three cases where a summary point is not mentioned ie First if no 1s are present we treat as 1 since
not mentioning a negative summary means a positive opinion Else if no 0s are present we treat as 1 Else when
both 1 and 0 are present some ground truth text have an explic it positive opinion some others have an explicit
negative opinion We treat as 1 an implicit positive opinion since a negative opinion is not mentioned
11ProofOAperfectly processes the text report of the agent to obtain a n umerical report r
01m By Assumption 1 the properness of ElicitationGPTfollows from the properness of
knowitornot scoring rules
We introduce Theorem 2 for imperfect oracles When the oracl es make errors the properness of
ElicitationGPTmay be broken For example if the question answering oracle OAﬂips the answers
on particular queries it is unclear whether ElicitationGPTis still proper Theorem 2 guarantees
the nonmanipulability when the agent does not have informa tion about the text
Theorem 2 For ElicitationGPTwith imperfect language oracles the optimal score of an age nt
who does not look at the submission and with any ﬁxed adversar ial manipulation is the prior score
Proof Since the agent does not exert any eﬀort ie has not looked at the homework submission

While peers can leave a review only on submission length with out much eﬀort the score given by the

instructor usually does not depend on the cheap signal of len gth Thus the states θ θ1θm

are not equally important for scoring textual report qualit y and the less important ones need to

be ﬁltered

Our ﬁltering method identiﬁes important summary points for scoring by partitioning the states

by semantic meanings into topics t t1tz For example three states θ1θ2θ3for correct

hypothesis base case and induction step in proof can be gro uped into topic t1for the proof

correctness State θ4for typos can be grouped into topic t2of the overall clarity Intuitively

the number of diﬀerent indicator states within a topic is posi tively related to a high cost to obtain

information on that topic The ﬁltering approach ignores re ports on the topics with fewer summary

points instead scoring only the top 2 topics with the most su mmary points ie the most dimensions

to report

The ﬁltered average aggregation AF with top 2 topics is deﬁ ned as the following

Deﬁnition 9 Filtered Average AF Given scoring rules S1S z andi1andi2as the indices

of the remaining topics after ﬁltering A ﬁltered average ag gregation outputs S1

2Si1Si2

In the deﬁnition zis the total number of topics and S1S zare multidimensional scoring

rules that can be applied to multiple indicator states in eac h topic

9222 Textual Scoring Rules We Test

For textual scoring we use the Vshaped single dimensional scoring rules for each indicator state

on terminal nodes in the aggregation hierarchy

AV The average scoring rule of Vshaped on each indicator state

AMV The average of the maxoverseparate scoring rule applied t o each topic

AFV The average scoring rule of Vshaped ﬁltered to the top 2 dive rse topics

AFMV The average of the maxoverseparate scoring rule ﬁltered t o the top 2 diverse topics

3 ElicitationGPT Scoring Text with Oracle Access to LLM

In this section we present ElicitationGPT an implementation of scoring rules for text via queries

to large language models LLMs Instead of presenting deta ils of the prompts we model LLMs

as oracles with the fundamental functionalities of summarization andquestion answering We

present the algorithm via queries to these oracles This ora cle abstraction allows the main ideas

of the algorithm and its properties to be studied without det ails of prompt engineering tricks

in implementation For example the properness of the textu al elicitation mechanism is clearly

inherited from the proper scoring rule from which it is const ructed

In our oracle model the LLM is able to accept queries to two fu ndamental tasks In the

literature of natural language processing these tasks are also fundamental benchmarks to evaluate

LLM performance We refer to oracle queries as OSandOAas deﬁned below

Summarization OS

Input a list Iiiof documents

Output a list T1T2Tm of summary points

Our deﬁnition of the summarization task is a variant of the cl assic summarization task and

it is also known as the key point summarization task BarHai m et al 2020

Question Answering OA

Input text review Rand a list of text statements T1Tm

Output a vector r 01mof summarization points on T1Tm On each dimension

ri 1 ifRhas a segment supporting a positive opinion on Ti 0 if a negative opinion or if

Rdoes not mention anything related to Ti

The query for OAis DoesRsupport a positive opinion on the following statements T1Tm

Datasets for question answering include Rajpurkar et al 2 016 Kwiatkowski et al 2019

etc

Deﬁnition 10 ElicitationGPTis deﬁned from a large language model that gives summarizati on

and question answering oracles OSandOAand an knowitornot scoring rule S

Input

a list of ground truth text I1Iswithin the same cluster eg instructor reviews on

submissions to the same homework assignment

10an agent report Pcorresponding to the kth ground truth eg a peer review on the same

submission as the kth instructor review

Output A score in 01on the agent report

1extracting dimensions of summary points

Query summarization oracle and get a list of summary points f rom the reviews

T1Tm OSI1Is

2clustering topics optional

Cluster the summary points by similar topics Required for AF MV and AFV scoring rule

Topics t1tz OST1Tm

3calculating prior

Process the ground truths within the same cluster to determi ne the prior distribution over the

semantic state

For each ground truth text Ii determine the indicator state for each summary point

θiOAIiT1Tm

For each indicator state θj count the frequency pθjof1s from ground truth2

4mapping agent report to belief Construct the mapping from the agent report to beliefs on

the true semantic state

rOART1Tm

5knowitornot scoring rule Apply the scoring rule Sprθk

31 Properness of ElicitationGPT

An important property of numerical scoring rules is propern ess ie incentivizing the forecaster to

report their true beliefs In this section we give two result s about the incentives of ElicitationGPT

First we consider the perfect language oracle model and sh ow that in this model ElicitationGPTis

proper ie the agents optimal strategy is to report thei r true belief about how the instructor would

review the submission Second we show that if the agent does not exert eﬀort ie in peer grading

does not look at the submission then their score for any pos sible adversarial manipulation and

for imperfect language oracles is upper bounded by the scor e from reporting the prior mean

which they can easily obtain by reporting I dont know

Theorem 1 ElicitationGPTwith perfect language oracles is proper for knowitornot beliefs ie

the optimal strategy of an agent who knows the prior and has kn owitornot beliefs is to report an

accurate textual representation of their belief

2We mark 1 as a positive opinion on a summary point For each ind icator states and for ground truth text there

are three cases where a summary point is not mentioned ie First if no 1s are present we treat as 1 since

not mentioning a negative summary means a positive opinion Else if no 0s are present we treat as 1 Else when

both 1 and 0 are present some ground truth text have an explic it positive opinion some others have an explicit

negative opinion We treat as 1 an implicit positive opinion since a negative opinion is not mentioned

11ProofOAperfectly processes the text report of the agent to obtain a n umerical report r

01m By Assumption 1 the properness of ElicitationGPTfollows from the properness of

knowitornot scoring rules

We introduce Theorem 2 for imperfect oracles When the oracl es make errors the properness of

ElicitationGPTmay be broken For example if the question answering oracle OAﬂips the answers

on particular queries it is unclear whether ElicitationGPTis still proper Theorem 2 guarantees

the nonmanipulability when the agent does not have informa tion about the text

Theorem 2 For ElicitationGPTwith imperfect language oracles the optimal score of an age nt

who does not look at the submission and with any ﬁxed adversar ial manipulation is the prior score

Proof Since the agent does not exert any eﬀort ie has not looked at the homework submission