tensorkelechi/arxiv-cstext · Datasets at Hugging Face

text stringlengths 0 89.3k
truth answers each on a multipoint scale In the latter th e model is given only the student answer
and ground truth answer not the question They ﬁnd that the re is little Pearson correlation
between instructor grades of the student answers and the clo seness reported by the language model
of the student and instructor answers In contrast to this ap proach our work constructs a multi
dimensional scoring rule from an analysis of several instru ctor answers ie instructor reviews to
similar questions ie submissions of the same assignmen t and then evaluates a students answer
ie peer review according to the surprising dimensions We have favorable results presented
subsequently
Characterization of Proper Scoring Rules There is an extensive literature that develops
numerical proper scoring rules where forecasters are incentivized to report their true bel iefs
4These proper scoring rules are characterized for eliciting beliefs over multidimensional states by
McCarthy 1956 and for multidimensional elicitation of t he mean by Abernethy and Frongillo
2012 Lambert 2011 characterized statistics that are d irectly elicitable
Optimization of Scoring Rules Surprisingly until recently there was relatively little w ork
on optimizing scoring rules subject to properness and boun dedness Motivated by peer grad
ing Li Hartline Shan and Wu 2022 optimizes scoring rul es for binary eﬀort where a peer ei
ther with no eﬀort reports a prior belief or exerts costly eﬀo rt to obtain and report a pos
terior belief Hartline Shan Li and Wu 2023 generalize the model to allow multidimensional
eﬀort as a knapsack problem for scoring rule optimization Ou r paper adapts the scoring rules
identiﬁed by these papers to scoring text Additional work o n scoring rule optimization includes
Papireddygari and Waggoner 2022 which considers connec tions between optimizing scoring rules
and contract theory and Chen and Yu 2021 which relaxes th e assumption that the prior is known
to the designer
Human Computation A central question in the ﬁeld of human computation is how alg orithms
can solve complex problems by organizing them into small tas ks that are assigned to humans For
example Soylent is a Microsoft Word plugin that can be used t o conduct standard editing tasks
on text documents Bernstein et al 2010 Tasks it support s include proofreading and shortening
In this computational model humans are used to perform stra ightforward comparisons between
text segments or generations of short text segments They ar e not asked for domain knowledge A
key concern in such systems is how to get reliable results whe n individual human answers may be
unreliable For example Soylent employs a ﬁndﬁxverify p aradigm that collects responses from
humans to ﬁndplaces in the text that could be shortened ﬁxthe text by shortening it and verify
that the shortened text have the same meaning For additiona l background see the short survey of
Miller et al 2010 or book of Law and Von Ahn 2011 Our meth od of constructing a scoring rule
for text via oracle calls to answer simple domainfree quest ions about text parallels the canonical
model of human computation Unlike the models of human compu tation we ﬁnd for our task of
grading peer reviews that prompting a language model for te xt analysis and comparisons performs
well without building in explicit methods for error detecti on and correction
Mechanism Design with LLMs There are other potential connections between mechanism
design and large language models For example Duetting et a l 2023 consider a setting where
agents are in competition to create some text like companie s creating a shared advertisement and
each agent prefers the outcome of their own language model T hey deﬁne a token auction model
and within it a secondprice auction
2 Model and Preliminaries
In this section we introduce the deﬁnition of proper scorin g rules We start deﬁning a scoring rule
for numerical reports in the classic model of information el icitation then proceed to the special case
of text elicitation We illustrate our deﬁnitions with exam ples from the peer grading application
521 Numerical Mean Elicitation
The principal mechanism designer posts a list of nexplicit dimensions for the agent to report
a numerical prediction The principal elicits agents repo rt on the multidimensional state θ
Θ 01n For example in peer grading the rubric consists of Statem ent of Result Proof and
Clarity which are dimensions for assessment of the homewor k quality 1 is the best quality on that
dimension The agent holds a multidimensional private bel iefqΘ about the states Let
µq01nbe the marginal means of the belief space The principal is in terested in eliciting the
marginal means of the agents private belief ie the agent only needs to report a single real number
for each dimension The report space Ris thus the same 0 1nas the state space
Before reporting the agent holds prior belief p01n about the states and learns by
receiving signal sScorrelated with the random state An information structure is a joint
distribution Θ S Upon receiving signal sand Bayesian updating the agent holds posterior
beliefqs Prθs on the state
The agent is scored by comparing the report r01nwith the ground truth state θ01n
The literature McCarthy 1956 Gneiting 2011 focuses on the design of proper scoring rules
which elicit truthful reports from the agent From the agent s perspective a scoring rule is proper
if reporting their true belief gains a weakly higher expec ted score than any other reports By
deﬁnition the report that maximizes the score is the Bayesi an optimal report
Deﬁnition 1 Properness A scoring rule SRΘRis proper for mean elicitation if for
any private belief qof the agent with mean µq and any deviation report rR
EθqSµqθEθqSrθ
In this paper we test multidimensional scoring rules ie scoring rules for multidimensional
reports Our multidimensional scoring rules can be decom posed into singledimensional scoring
rules Section 211 and a multidimensional aggregation rule Section 212
211 Singledimensional Scoring Rules
We introduce the singledimensional quadratic scoring rul e and the Vshaped scoring rule Li et al
2022 in this section We note that the quadratic scoring rul e is only used for numeric reviews in
our experimental comparison while Vshaped scoring rule i s used for both numerical reviews and
textual reviews
Deﬁnition 2 Quadratic A quadratic scoring rule is Srθ 1rθ2rθ01
The Vshaped scoring rule partitions the report space into a trinary space a report higher than
prior lower than prior or the same as prior µp Higher or lower than prior are cases when the
agent has information ie signal about the state θ whereas reporting prior can be interpreted as
I dont know
Deﬁnition 3 Vshaped A Vshaped scoring rule S 010101for mean elicitation is
deﬁned with the prior mean µp01 Whenµp12
Sµprθ

341
2θµp
1µpifr µp
141
2θµp
1µpifr µp
12 else
601
12
1 0µpriorS10S00
S01S11
state report
score
Figure 1 The Vshaped scoring rule the optimal singledim ensional scoring rule from Li et al
2022 Once ﬁxing the report r the score is linear in the state θ The scoring rule oﬀers two linear

truth answers each on a multipoint scale In the latter th e model is given only the student answer

and ground truth answer not the question They ﬁnd that the re is little Pearson correlation

between instructor grades of the student answers and the clo seness reported by the language model

of the student and instructor answers In contrast to this ap proach our work constructs a multi

dimensional scoring rule from an analysis of several instru ctor answers ie instructor reviews to

similar questions ie submissions of the same assignmen t and then evaluates a students answer

ie peer review according to the surprising dimensions We have favorable results presented

subsequently

Characterization of Proper Scoring Rules There is an extensive literature that develops

numerical proper scoring rules where forecasters are incentivized to report their true bel iefs

4These proper scoring rules are characterized for eliciting beliefs over multidimensional states by

McCarthy 1956 and for multidimensional elicitation of t he mean by Abernethy and Frongillo

2012 Lambert 2011 characterized statistics that are d irectly elicitable

Optimization of Scoring Rules Surprisingly until recently there was relatively little w ork

on optimizing scoring rules subject to properness and boun dedness Motivated by peer grad

ing Li Hartline Shan and Wu 2022 optimizes scoring rul es for binary eﬀort where a peer ei

ther with no eﬀort reports a prior belief or exerts costly eﬀo rt to obtain and report a pos

terior belief Hartline Shan Li and Wu 2023 generalize the model to allow multidimensional

eﬀort as a knapsack problem for scoring rule optimization Ou r paper adapts the scoring rules

identiﬁed by these papers to scoring text Additional work o n scoring rule optimization includes

Papireddygari and Waggoner 2022 which considers connec tions between optimizing scoring rules

and contract theory and Chen and Yu 2021 which relaxes th e assumption that the prior is known

to the designer

Human Computation A central question in the ﬁeld of human computation is how alg orithms

can solve complex problems by organizing them into small tas ks that are assigned to humans For

example Soylent is a Microsoft Word plugin that can be used t o conduct standard editing tasks

on text documents Bernstein et al 2010 Tasks it support s include proofreading and shortening

In this computational model humans are used to perform stra ightforward comparisons between

text segments or generations of short text segments They ar e not asked for domain knowledge A

key concern in such systems is how to get reliable results whe n individual human answers may be

unreliable For example Soylent employs a ﬁndﬁxverify p aradigm that collects responses from

humans to ﬁndplaces in the text that could be shortened ﬁxthe text by shortening it and verify

that the shortened text have the same meaning For additiona l background see the short survey of

Miller et al 2010 or book of Law and Von Ahn 2011 Our meth od of constructing a scoring rule

for text via oracle calls to answer simple domainfree quest ions about text parallels the canonical

model of human computation Unlike the models of human compu tation we ﬁnd for our task of

grading peer reviews that prompting a language model for te xt analysis and comparisons performs

well without building in explicit methods for error detecti on and correction

Mechanism Design with LLMs There are other potential connections between mechanism

design and large language models For example Duetting et a l 2023 consider a setting where

agents are in competition to create some text like companie s creating a shared advertisement and

each agent prefers the outcome of their own language model T hey deﬁne a token auction model

and within it a secondprice auction

2 Model and Preliminaries

In this section we introduce the deﬁnition of proper scorin g rules We start deﬁning a scoring rule

for numerical reports in the classic model of information el icitation then proceed to the special case

of text elicitation We illustrate our deﬁnitions with exam ples from the peer grading application

521 Numerical Mean Elicitation

The principal mechanism designer posts a list of nexplicit dimensions for the agent to report

a numerical prediction The principal elicits agents repo rt on the multidimensional state θ

Θ 01n For example in peer grading the rubric consists of Statem ent of Result Proof and

Clarity which are dimensions for assessment of the homewor k quality 1 is the best quality on that

dimension The agent holds a multidimensional private bel iefqΘ about the states Let

µq01nbe the marginal means of the belief space The principal is in terested in eliciting the

marginal means of the agents private belief ie the agent only needs to report a single real number

for each dimension The report space Ris thus the same 0 1nas the state space

Before reporting the agent holds prior belief p01n about the states and learns by

receiving signal sScorrelated with the random state An information structure is a joint

distribution Θ S Upon receiving signal sand Bayesian updating the agent holds posterior

beliefqs Prθs on the state

The agent is scored by comparing the report r01nwith the ground truth state θ01n

The literature McCarthy 1956 Gneiting 2011 focuses on the design of proper scoring rules

which elicit truthful reports from the agent From the agent s perspective a scoring rule is proper

if reporting their true belief gains a weakly higher expec ted score than any other reports By

deﬁnition the report that maximizes the score is the Bayesi an optimal report

Deﬁnition 1 Properness A scoring rule SRΘRis proper for mean elicitation if for

any private belief qof the agent with mean µq and any deviation report rR

EθqSµqθEθqSrθ

In this paper we test multidimensional scoring rules ie scoring rules for multidimensional

reports Our multidimensional scoring rules can be decom posed into singledimensional scoring

rules Section 211 and a multidimensional aggregation rule Section 212

211 Singledimensional Scoring Rules

We introduce the singledimensional quadratic scoring rul e and the Vshaped scoring rule Li et al

2022 in this section We note that the quadratic scoring rul e is only used for numeric reviews in

our experimental comparison while Vshaped scoring rule i s used for both numerical reviews and

textual reviews

Deﬁnition 2 Quadratic A quadratic scoring rule is Srθ 1rθ2rθ01

The Vshaped scoring rule partitions the report space into a trinary space a report higher than

prior lower than prior or the same as prior µp Higher or lower than prior are cases when the

agent has information ie signal about the state θ whereas reporting prior can be interpreted as

I dont know

Deﬁnition 3 Vshaped A Vshaped scoring rule S 010101for mean elicitation is

deﬁned with the prior mean µp01 Whenµp12

Sµprθ

341

2θµp

1µpifr µp

141

2θµp

1µpifr µp

12 else

601

12

1 0µpriorS10S00

S01S11

state report

score

Figure 1 The Vshaped scoring rule the optimal singledim ensional scoring rule from Li et al

2022 Once ﬁxing the report r the score is linear in the state θ The scoring rule oﬀers two linear