tensorkelechi/arxiv-cstext · Datasets at Hugging Face

text stringlengths 0 89.3k
score functions for the agent to select When rµp the agent selects the line S00 toS01
Otherwise the agent selects the line S10 toS11
Whenµp12 the Vshaped scoring rule is Sµprθ S1µp1r1θ
The Vshaped scoring rule can be equivalently implemented a s asking the agent to report if the
mean of his belief is higher or lower than the prior mean µp Figure 1 geometrically explains the
Vshaped scoring rule Fixing report r the score is linear in state θ The Vshaped scoring rule
gives the lowest expected score 12on prior report a high expost score on a surprisingly corre ct
report the right half of the thick line and a low expost sc ore on a surprisingly incorrect report
the right half of the thin line The side that the prior pred icts to be less often realized is the
surprising side
212 Multidimensional Aggregations of Scoring Rules
An aggregation of scoring rules takes a set of scoring rules S1S n reports r1rn and a
multidimensional state θ1θnas input and outputs a realvalued score
In this paper we are interested in two basic multidimensio nal aggregations of proper scoring
rules for mean elicitation the average aggregation and the maxoverseparate aggregation These
basic aggregation methods can also be combined hierarchica lly
The average aggregation is deﬁned as the the average of input scores and is a standard approach
in theory and practice
Deﬁnition 4 Average Aggregation Given scoring rules S1S n an average scoring rule Sis
S1
nsummationtextn
i1Si
Introduced by Li et al 2022 the maxoverseparate aggre gation scores the agent on the di
mension for which the agent has highest expected score accor ding to their posterior belief The
maxoverseparate over Vshaped singledimensional scor e is shown to be approximately optimal for
incentivizing binary eﬀort and is unboundedly better than t he optimal scoring rule under average
aggregation
Deﬁnition 5 Maxoverseparate Aggregation Given scoring rules S1S n a maxoverseparate
aggregation SisSSiwherei arg maxiEθSi
These basis aggregation methods can be combined in a hierarc hy for example we will be eval
uating two level hierarchies that partition the dimensions and apply maxoverseparate in each
partition and then average across partitions Maxoverse parate is better for incentives while aver
age aggregation is less noisy hierarchical aggregation co mbines the good factors of both
7Deﬁnition 6 Aggregation Hierarchy A scoring rule with aggregation hierarchy is represented as
a tree In the tree all terminal nodes on the tree have the same depths from the root node Each
terminal node corresponds to a singledimensional scoring rule Vshaped or quadratic On each
level each node represents the same aggregation which tak es input a set of scoring rules S1S z
in its child nodes and outputs one score
The scoring rule with an aggregation hierarchy is named afte r the aggregation methods with
A for the average aggregation and M for the maxoversepara te aggregation For example AV
stands for average aggregation A over Vshaped singledi mensional scores V AMV stands for
an aggregation hierarchy with two levels where the termina l nodes are Vshaped singledimensional
scores V grouped and connected to parent nodes for maxov erseparate aggregation M and the
root is the average aggregation A over the maxoversepar ate scores
22 Textual Elicitation
Unlike explicitly given dimensions for numerical reports a textual report consists of implicit indi
cators for summary points Each textual ground truth indica tes the presence of mbinary states
in a vector θ θ1θ2θm Each state θi 01indicates agree1 or disagree0 on a
summary point For example in a peer review of an induction h omework θ1can be whether the
hypothesis is correctly stated θ2for whether the base case is correct θ3for whether the induction
step is correct and θ4for the nonexistence of typos etc A textual report can als o havefor
not applicable on a summary point Thus the report space i sri 01
We make the knowitornot assumption on the information st ructure that the agent either
knows the truth or does not know anything In the former case where the agent knows the truth
the agents belief is deterministically some state 0 or 1 In the later case where the agent does not
learn anything the belief is the same as prior
Assumption 1 Knowitornot On each dimension iof the state θ the agents posterior distri
butionqiis either the truth or the prior pθi ieqi 01pθi
Assumption 1 induces special case of proper scoring rules fo r knowitornot indicators Given
any proper scoring rule for probabilistic belief we can deﬁ ne a proper scoring rule for knowitor
not indicators which ﬁrst maps a report r 01mto a probabilistic belief m
i101pθi
then applies a proper scoring rule to the probabilistic beli ef
Deﬁnition 7 Proper Scoring Rules for Knowitornot Indicators1Fixing prior p pθiion
binary indicators a scoring rule Sp01ℓ 01ℓ01for knowitornot indicators is
proper if there exists a proper scoring rule S 01ℓ01ℓ01on belief space that
Sprθ Srrθ
1There exists an alternative deﬁnition of properness for kno witornot indicators Given the same mapping rr
from report to the probabilistic belief a scoring rule for k nowitornot indicators is proper if EθrrSprθ
EθrrSprθr 01ℓ This alternative deﬁnition is similar as the deﬁnition of p roper scoring rule
for general beliefs and is equivalent to Deﬁnition 7 It is s traightforward to see that Deﬁnition 7 satisﬁes the
requirements in the alternative deﬁnition To see the alter native deﬁnition also satisﬁed Deﬁnition 7 we can construc t
a proper scoring rule Sfor general beliefs from a proper scoring rule Spfor knowitornot indicators Srθ
argmaxrEθrSprθ where ris the probabilistic belief
8where the rrmaps report to the probabilistic belief
ribraceleftbiggriifri 01
pθielseri
For a single indicator state any scoring rules look similar to a Vshaped scoring rule since
the report space is trinary We rewrite the single dimension al Vshaped scoring rules for knowit
ornot indicators as a building block of multidimensional knowitornot scoring rules Recall the
Vshaped scoring rule is parameterized to have the tip of the V at the prior p
Deﬁnition 8 Vshaped for Knowitornot Indicators Under Assumption 1 a Vshaped single
dimensional scoring rule is S01 01 R When prior p Prθ 112is leaning
towards 0
Sprθ

1 ifrθ 1
0ifr 1θ 0
05 p
21pifrθ 0
05p
21pifr 0θ 1
05 ifr1
When prior p 12is leaning towards 1Srθ S1p1r1θ
221 Additional Aggregation for Text Filtered Average Aggregation
In addition to the aggregation method in Section 212 we in troduce the ﬁltered average aggregation
speciﬁcally for text The ﬁltered average aggregation skip s dimensions of cheap signals and scores
only a subset of summary points The cheap signal problem com es from the highdimensional
nature of summary points in text Brooks et al 2022 Kong an d Schoenebeck 2018 For example
a review segment commenting on clarity or length is less impo rtant than on correctness of proof

score functions for the agent to select When rµp the agent selects the line S00 toS01

Otherwise the agent selects the line S10 toS11

Whenµp12 the Vshaped scoring rule is Sµprθ S1µp1r1θ

The Vshaped scoring rule can be equivalently implemented a s asking the agent to report if the

mean of his belief is higher or lower than the prior mean µp Figure 1 geometrically explains the

Vshaped scoring rule Fixing report r the score is linear in state θ The Vshaped scoring rule

gives the lowest expected score 12on prior report a high expost score on a surprisingly corre ct

report the right half of the thick line and a low expost sc ore on a surprisingly incorrect report

the right half of the thin line The side that the prior pred icts to be less often realized is the

surprising side

212 Multidimensional Aggregations of Scoring Rules

An aggregation of scoring rules takes a set of scoring rules S1S n reports r1rn and a

multidimensional state θ1θnas input and outputs a realvalued score

In this paper we are interested in two basic multidimensio nal aggregations of proper scoring

rules for mean elicitation the average aggregation and the maxoverseparate aggregation These

basic aggregation methods can also be combined hierarchica lly

The average aggregation is deﬁned as the the average of input scores and is a standard approach

in theory and practice

Deﬁnition 4 Average Aggregation Given scoring rules S1S n an average scoring rule Sis

S1

nsummationtextn

i1Si

Introduced by Li et al 2022 the maxoverseparate aggre gation scores the agent on the di

mension for which the agent has highest expected score accor ding to their posterior belief The

maxoverseparate over Vshaped singledimensional scor e is shown to be approximately optimal for

incentivizing binary eﬀort and is unboundedly better than t he optimal scoring rule under average

aggregation

Deﬁnition 5 Maxoverseparate Aggregation Given scoring rules S1S n a maxoverseparate

aggregation SisSSiwherei arg maxiEθSi

These basis aggregation methods can be combined in a hierarc hy for example we will be eval

uating two level hierarchies that partition the dimensions and apply maxoverseparate in each

partition and then average across partitions Maxoverse parate is better for incentives while aver

age aggregation is less noisy hierarchical aggregation co mbines the good factors of both

7Deﬁnition 6 Aggregation Hierarchy A scoring rule with aggregation hierarchy is represented as

a tree In the tree all terminal nodes on the tree have the same depths from the root node Each

terminal node corresponds to a singledimensional scoring rule Vshaped or quadratic On each

level each node represents the same aggregation which tak es input a set of scoring rules S1S z

in its child nodes and outputs one score

The scoring rule with an aggregation hierarchy is named afte r the aggregation methods with

A for the average aggregation and M for the maxoversepara te aggregation For example AV

stands for average aggregation A over Vshaped singledi mensional scores V AMV stands for

an aggregation hierarchy with two levels where the termina l nodes are Vshaped singledimensional

scores V grouped and connected to parent nodes for maxov erseparate aggregation M and the

root is the average aggregation A over the maxoversepar ate scores

22 Textual Elicitation

Unlike explicitly given dimensions for numerical reports a textual report consists of implicit indi

cators for summary points Each textual ground truth indica tes the presence of mbinary states

in a vector θ θ1θ2θm Each state θi 01indicates agree1 or disagree0 on a

summary point For example in a peer review of an induction h omework θ1can be whether the

hypothesis is correctly stated θ2for whether the base case is correct θ3for whether the induction

step is correct and θ4for the nonexistence of typos etc A textual report can als o havefor

not applicable on a summary point Thus the report space i sri 01

We make the knowitornot assumption on the information st ructure that the agent either

knows the truth or does not know anything In the former case where the agent knows the truth

the agents belief is deterministically some state 0 or 1 In the later case where the agent does not

learn anything the belief is the same as prior

Assumption 1 Knowitornot On each dimension iof the state θ the agents posterior distri

butionqiis either the truth or the prior pθi ieqi 01pθi

Assumption 1 induces special case of proper scoring rules fo r knowitornot indicators Given

any proper scoring rule for probabilistic belief we can deﬁ ne a proper scoring rule for knowitor

not indicators which ﬁrst maps a report r 01mto a probabilistic belief m

i101pθi

then applies a proper scoring rule to the probabilistic beli ef

Deﬁnition 7 Proper Scoring Rules for Knowitornot Indicators1Fixing prior p pθiion

binary indicators a scoring rule Sp01ℓ 01ℓ01for knowitornot indicators is

proper if there exists a proper scoring rule S 01ℓ01ℓ01on belief space that

Sprθ Srrθ

1There exists an alternative deﬁnition of properness for kno witornot indicators Given the same mapping rr

from report to the probabilistic belief a scoring rule for k nowitornot indicators is proper if EθrrSprθ

EθrrSprθr 01ℓ This alternative deﬁnition is similar as the deﬁnition of p roper scoring rule

for general beliefs and is equivalent to Deﬁnition 7 It is s traightforward to see that Deﬁnition 7 satisﬁes the

requirements in the alternative deﬁnition To see the alter native deﬁnition also satisﬁed Deﬁnition 7 we can construc t

a proper scoring rule Sfor general beliefs from a proper scoring rule Spfor knowitornot indicators Srθ

argmaxrEθrSprθ where ris the probabilistic belief

8where the rrmaps report to the probabilistic belief

ribraceleftbiggriifri 01

pθielseri

For a single indicator state any scoring rules look similar to a Vshaped scoring rule since

the report space is trinary We rewrite the single dimension al Vshaped scoring rules for knowit

ornot indicators as a building block of multidimensional knowitornot scoring rules Recall the

Vshaped scoring rule is parameterized to have the tip of the V at the prior p

Deﬁnition 8 Vshaped for Knowitornot Indicators Under Assumption 1 a Vshaped single

dimensional scoring rule is S01 01 R When prior p Prθ 112is leaning

towards 0

Sprθ

1 ifrθ 1

0ifr 1θ 0

05 p

21pifrθ 0

05p

21pifr 0θ 1

05 ifr1

When prior p 12is leaning towards 1Srθ S1p1r1θ

221 Additional Aggregation for Text Filtered Average Aggregation

In addition to the aggregation method in Section 212 we in troduce the ﬁltered average aggregation

speciﬁcally for text The ﬁltered average aggregation skip s dimensions of cheap signals and scores

only a subset of summary points The cheap signal problem com es from the highdimensional

nature of summary points in text Brooks et al 2022 Kong an d Schoenebeck 2018 For example

a review segment commenting on clarity or length is less impo rtant than on correctness of proof