tensorkelechi/arxiv-cstext · Datasets at Hugging Face

text stringlengths 0 89.3k
that may depend on the prior distribution
The natural language processing tasks in the above framewor k can be implemented with domain
knowledgefree queries to a language model Speciﬁcally t he summary points can be identiﬁed
with a summarization oracle and the response text can be mapp ed into response vectors with a
questionanswering oracle These tasks are standard in nat ural language processing
Textual scoring rules constructed from the framework inher it the properness property of the
proper scoring rule of step 4 above Properness implies tha t reporting the ground truth including
uncertainty on the truth gains the highest expected score
Many scoring rules can be applied to the multidimensional s pace of summary points These
scoring rules are speciﬁed by a singledimensional scoring rule and a method of aggregation For
example a classical singledimensional scoring rule is th e quadratic scoring rule and a classical
method of aggregation is averaging Another example in Li et al 2022 deﬁnes approximately
optimal scoring rules for incentivizing binary eﬀort using a singledimensional Vshaped scoring
rule and aggregating by taking the maximum according to the b elief of the report which for
knowitornot beliefs corresponds to the one where the kno wledge is furthest from the prior Our
empirical analysis considers these and related scoring rul es
This paper aims to evaluate proper scoring rules for alignme nt with humans We empirically
evaluate the proposed scoring rules on a peergrading data s et In peer grading student submissions
are graded by their peers See a detailed introduction to thi s application in Section 11 This dataset
contains the following
textual and numeric peer reviews and instructor reviews of m ultiple submissions for multiple
assignments
instructor scores of the peer reviews and
overall grades for the students in the class including home works peer reviews and exams
2We make a simplifying assumption that the information sourc e has knowitornot beliefs mean
ing they either know the ground truth or their belief is the s ame as prior ie knowing nothing
This assumption is appropriate in settings where the belief comes from a signal that provides
evidence and fully reveals the ground truth The peer gradin g data is consistent with our assump
tion The points reported in a peers review are supported wi th evidence from the homework The
assumption also restricts a report with uncertainty to sayi ng I dont know or omitting discussion
Our main analysis is of the alignment measured as rank corre lation of the textual scoring
rule applied to the peer reviews with ground truth review giv en by the instructor scores of the
peer reviews This analysis shows a highdegree of alignmen t and that the scoring rules for text
are better aligned than traditional numeric scoring rules applied to the numerical peer reviews
To evaluate the possibility that scoring rules for text migh t be more reliable for evaluating the
peer reviews than the human instructor we compare these s cores for alignment with the overall
student grades We ﬁnd that the text scoring rules are more al igned with overall student grades
than the instructors scores
11 Introduction to Peer Grading
A central application for the methods of the paper is peer gra ding In a course that uses peer
grading the students turn in submissions for a number of hom ework assignments and the students
in this role called peers also review each others submiss ions both quantitatively by providing
numeric scores and qualitatively by providing textual fe edback according to a multidimensional
rubric For an algorithms course the rubric may consist of t he algorithm the analysis and
writing quality
There are several beneﬁts of peer grading First peer gradi ng improves the learning outcomes
of students Reading peers homework submission helps stud ent learn from diﬀerent perspectives
as well as the mistakes and successes from their peers Secon d peer grading eﬀectively scales the
grading process Distributing the grading tasks among stud ents alleviates the workload of the
instructor which is especially helpful in a large class Ho wever peer grading also creates new
grading tasks as the peer reviews like any assigned work mu st be graded
To alleviate the additional grading burden prior work of Li et al 2022 developed numerical
scoring rules for the peer grading application It is more im pactful however to grade the written
feedback in peer reviews than it is to grade the numerical sco res First it places the emphasis of
the peer review task on giving good written feedback where a nswering openended review ques
tions improves learning more than ﬁnegrained numerical gr ading tasks Second it can potentially
be done more accurately While the peer and instructor may di sagree on the score to assign a
submission it is easier to agree more qualitatively on what the submission does well and where
it has mistakes or needs improvement Thus generalizing sc oring rules for numbers to scoring
rules for text has the potential both to emphasize the right a ctivities and to be more accurate in
the assessment of the peer reviews Note that the developmen t of these scoring rules is critical to
scaling of large courses via peer grading without increasin g the grading workload of the instructor
The peer grading application has special structure that fac ilitates the design of scoring rules
for text As discussed in the introduction the framework fo r textual reviews assumes that the
prompts corresponding to homework submissions and groun d truth responses corresponding to
instructor reviews are partitioned into clusters for the purpose of calculating the prior distribution
of the ground truth The peer grading has a natural partitio ning to clusters based on homework
problems Speciﬁcally for each homework problem the inst ructor grades multiple submissions
3These submissions and the instructor review form a cluster The prior distribution of the ground
truth for such a cluster can be interpreted for example as s pecifying the frequency of mistakes in
the submissions for this problem This clustering allows th e distribution of mistakes to be diﬀerent
for diﬀerent homework problems For example in an algorithm s class submissions to a dynamic
programming problem and a proof by induction will have a diﬀer ent distribution of mistakes
12 Related Work
Textual Elicitation Motivated by the recent development of language models sev eral papers
aim to design scoring mechanism to evaluate models truthful ly Kimpara et al 2023 view a lan
guage model as producing a distribution over responses and d esign scoring mechanisms for evalu
ating this distribution from independent samples While th e scoring mechanism in Kimpara et al
2023 evaluate the quality of the distribution where the te xtual report is drawn our scoring rule
evaluates the quality of the textual response itself where uncertainty can be expressed directly in
the text Independently and concurrently Lu et al 2024 c onsider peer prediction with textual
reports where responses are evaluated in comparison to pee r responses rather than ground truth
responses They use a pretrained language model to interpr et text as a probabilistic report of
peer responses and use a proper scoring rule to evaluate aga inst the peer responses instead of the
ground truth The goal in their paper is to distinguish betwe en GPT generated reviews and hu
man written reviews where they ﬁnd predicting the next word is better than scoring the semantic
meaning However our goal is to align with human preference where scoring next word has been
known to be bad on outofsample report such as I dont know Moreover there are two reasons
why their in their evaluation of scoring semantic meaning ca n be bad First they directly ask GPT
to interpret text as probabilities of semantic meaning whe re inaccurate GPT prior signiﬁcantly
reduces the performance Instead our approach of dataset p artitioning allows us to access accurate
prior Second they only test the log scoring rule while our paper compares diﬀerent scoring rules
and shows applying ﬁltering to semantic meaning signiﬁcant ly improves the alignment performance
Grading with LLMs Language models have very recently been studied for their us e in grading
textual responses of students Gao et al 2023 compare sev eral language models using an ap
proach based on tokenizing both student answers and ground t ruth and then comparing the vector
representations with cosine similarity They show that the approach works well for binary grad
ing of simple questions quiz questions but not as well on mu ltipoint grading of questions with
more complex answers activity questions Schneider et al 2023 consider prompting the language
model to both evaluate the quality of answers directly and co mpare answers of students to ground

that may depend on the prior distribution

The natural language processing tasks in the above framewor k can be implemented with domain

knowledgefree queries to a language model Speciﬁcally t he summary points can be identiﬁed

with a summarization oracle and the response text can be mapp ed into response vectors with a

questionanswering oracle These tasks are standard in nat ural language processing

Textual scoring rules constructed from the framework inher it the properness property of the

proper scoring rule of step 4 above Properness implies tha t reporting the ground truth including

uncertainty on the truth gains the highest expected score

Many scoring rules can be applied to the multidimensional s pace of summary points These

scoring rules are speciﬁed by a singledimensional scoring rule and a method of aggregation For

example a classical singledimensional scoring rule is th e quadratic scoring rule and a classical

method of aggregation is averaging Another example in Li et al 2022 deﬁnes approximately

optimal scoring rules for incentivizing binary eﬀort using a singledimensional Vshaped scoring

rule and aggregating by taking the maximum according to the b elief of the report which for

knowitornot beliefs corresponds to the one where the kno wledge is furthest from the prior Our

empirical analysis considers these and related scoring rul es

This paper aims to evaluate proper scoring rules for alignme nt with humans We empirically

evaluate the proposed scoring rules on a peergrading data s et In peer grading student submissions

are graded by their peers See a detailed introduction to thi s application in Section 11 This dataset

contains the following

textual and numeric peer reviews and instructor reviews of m ultiple submissions for multiple

assignments

instructor scores of the peer reviews and

overall grades for the students in the class including home works peer reviews and exams

2We make a simplifying assumption that the information sourc e has knowitornot beliefs mean

ing they either know the ground truth or their belief is the s ame as prior ie knowing nothing

This assumption is appropriate in settings where the belief comes from a signal that provides

evidence and fully reveals the ground truth The peer gradin g data is consistent with our assump

tion The points reported in a peers review are supported wi th evidence from the homework The

assumption also restricts a report with uncertainty to sayi ng I dont know or omitting discussion

Our main analysis is of the alignment measured as rank corre lation of the textual scoring

rule applied to the peer reviews with ground truth review giv en by the instructor scores of the

peer reviews This analysis shows a highdegree of alignmen t and that the scoring rules for text

are better aligned than traditional numeric scoring rules applied to the numerical peer reviews

To evaluate the possibility that scoring rules for text migh t be more reliable for evaluating the

peer reviews than the human instructor we compare these s cores for alignment with the overall

student grades We ﬁnd that the text scoring rules are more al igned with overall student grades

than the instructors scores

11 Introduction to Peer Grading

A central application for the methods of the paper is peer gra ding In a course that uses peer

grading the students turn in submissions for a number of hom ework assignments and the students

in this role called peers also review each others submiss ions both quantitatively by providing

numeric scores and qualitatively by providing textual fe edback according to a multidimensional

rubric For an algorithms course the rubric may consist of t he algorithm the analysis and

writing quality

There are several beneﬁts of peer grading First peer gradi ng improves the learning outcomes

of students Reading peers homework submission helps stud ent learn from diﬀerent perspectives

as well as the mistakes and successes from their peers Secon d peer grading eﬀectively scales the

grading process Distributing the grading tasks among stud ents alleviates the workload of the

instructor which is especially helpful in a large class Ho wever peer grading also creates new

grading tasks as the peer reviews like any assigned work mu st be graded

To alleviate the additional grading burden prior work of Li et al 2022 developed numerical

scoring rules for the peer grading application It is more im pactful however to grade the written

feedback in peer reviews than it is to grade the numerical sco res First it places the emphasis of

the peer review task on giving good written feedback where a nswering openended review ques

tions improves learning more than ﬁnegrained numerical gr ading tasks Second it can potentially

be done more accurately While the peer and instructor may di sagree on the score to assign a

submission it is easier to agree more qualitatively on what the submission does well and where

it has mistakes or needs improvement Thus generalizing sc oring rules for numbers to scoring

rules for text has the potential both to emphasize the right a ctivities and to be more accurate in

the assessment of the peer reviews Note that the developmen t of these scoring rules is critical to

scaling of large courses via peer grading without increasin g the grading workload of the instructor

The peer grading application has special structure that fac ilitates the design of scoring rules

for text As discussed in the introduction the framework fo r textual reviews assumes that the

prompts corresponding to homework submissions and groun d truth responses corresponding to

instructor reviews are partitioned into clusters for the purpose of calculating the prior distribution

of the ground truth The peer grading has a natural partitio ning to clusters based on homework

problems Speciﬁcally for each homework problem the inst ructor grades multiple submissions

3These submissions and the instructor review form a cluster The prior distribution of the ground

truth for such a cluster can be interpreted for example as s pecifying the frequency of mistakes in

the submissions for this problem This clustering allows th e distribution of mistakes to be diﬀerent

for diﬀerent homework problems For example in an algorithm s class submissions to a dynamic

programming problem and a proof by induction will have a diﬀer ent distribution of mistakes

12 Related Work

Textual Elicitation Motivated by the recent development of language models sev eral papers

aim to design scoring mechanism to evaluate models truthful ly Kimpara et al 2023 view a lan

guage model as producing a distribution over responses and d esign scoring mechanisms for evalu

ating this distribution from independent samples While th e scoring mechanism in Kimpara et al

2023 evaluate the quality of the distribution where the te xtual report is drawn our scoring rule

evaluates the quality of the textual response itself where uncertainty can be expressed directly in

the text Independently and concurrently Lu et al 2024 c onsider peer prediction with textual

reports where responses are evaluated in comparison to pee r responses rather than ground truth

responses They use a pretrained language model to interpr et text as a probabilistic report of

peer responses and use a proper scoring rule to evaluate aga inst the peer responses instead of the

ground truth The goal in their paper is to distinguish betwe en GPT generated reviews and hu

man written reviews where they ﬁnd predicting the next word is better than scoring the semantic

meaning However our goal is to align with human preference where scoring next word has been

known to be bad on outofsample report such as I dont know Moreover there are two reasons

why their in their evaluation of scoring semantic meaning ca n be bad First they directly ask GPT

to interpret text as probabilities of semantic meaning whe re inaccurate GPT prior signiﬁcantly

reduces the performance Instead our approach of dataset p artitioning allows us to access accurate

prior Second they only test the log scoring rule while our paper compares diﬀerent scoring rules

and shows applying ﬁltering to semantic meaning signiﬁcant ly improves the alignment performance

Grading with LLMs Language models have very recently been studied for their us e in grading

textual responses of students Gao et al 2023 compare sev eral language models using an ap

proach based on tokenizing both student answers and ground t ruth and then comparing the vector

representations with cosine similarity They show that the approach works well for binary grad

ing of simple questions quiz questions but not as well on mu ltipoint grading of questions with

more complex answers activity questions Schneider et al 2023 consider prompting the language

model to both evaluate the quality of answers directly and co mpare answers of students to ground