<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="shibata-miyamura-2026-preferences">
    <titleInfo>
        <title>From Preferences to Scores: Semi-Supervised Multi-Trait Essay Scoring with LLM Pairwise Preferences and Few Human Anchors</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Takumi</namePart>
        <namePart type="family">Shibata</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Yuichi</namePart>
        <namePart type="family">Miyamura</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <originInfo>
        <dateIssued>2026-11</dateIssued>
    </originInfo>
    <typeOfResource>text</typeOfResource>
    <relatedItem type="host">
        <titleInfo>
            <title>Proceedings of the 15th International Joint Conference on Natural Language Processing and the 5th Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics (Volume 1: Long Papers)</title>
        </titleInfo>
        <name type="personal">
            <namePart type="given">Yue</namePart>
            <namePart type="family">Zhang</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Ruifeng</namePart>
            <namePart type="family">Xu</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Ayu</namePart>
            <namePart type="family">Purwarianti</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Tong</namePart>
            <namePart type="family">Rong</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Jun</namePart>
            <namePart type="family">Suzuki</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <originInfo>
            <publisher>Association for Computational Linguistics</publisher>
            <place>
                <placeTerm type="text">Hengqin, China</placeTerm>
            </place>
        </originInfo>
        <genre authority="marcgt">conference publication</genre>
        <identifier type="isbn">979-8-89176-439-2</identifier>
    </relatedItem>
    <abstract>Multi-trait automated essay scoring (AES) provides trait-level scores as feedback but requires costly annotations across multiple traits to train accurate scoring models. To alleviate this annotation burden, we propose MTPAS (Multi-Trait Preference-to-Absolute Scoring), a semi-supervised framework that learns relative score orderings from LLM-generated pairwise preferences and calibrates them to absolute rubric scales using a small set of human-scored anchor essays in an end-to-end manner. To this end, MTPAS jointly optimizes three objectives: a Bradley–Terry pairwise preference loss, a regression loss on the anchor essays, and a Sinkhorn divergence-based distribution regularizer. The model adopts a bifactor architecture that decomposes essay representations into a shared general factor and trait-specific residuals, improving robustness against noisy LLM-generated preferences. We further introduce pairwise rubric compression, which reformulates the original score-level rubrics into concise comparative criteria, reducing the inference cost of collecting LLM preferences. On ASAP and ASAP++, MTPAS achieves 91.4% of the fully supervised SoTA performance with only 30 anchor essays per prompt and further narrows the gap with 200 anchors. Under the same anchor budget, MTPAS outperforms conventional multi-trait AES baselines at every anchor size investigated in this study.</abstract>
    <identifier type="citekey">shibata-miyamura-2026-preferences</identifier>
    <location>
        <url>https://aclanthology.org/2026.aacl-long.273/</url>
    </location>
    <part>
        <date>2026-11</date>
        <extent unit="page">
            <start>6041</start>
            <end>6057</end>
        </extent>
    </part>
</mods>
</modsCollection>

%0 Conference Proceedings
%T From Preferences to Scores: Semi-Supervised Multi-Trait Essay Scoring with LLM Pairwise Preferences and Few Human Anchors
%A Shibata, Takumi
%A Miyamura, Yuichi
%Y Zhang, Yue
%Y Xu, Ruifeng
%Y Purwarianti, Ayu
%Y Rong, Tong
%Y Suzuki, Jun
%S Proceedings of the 15th International Joint Conference on Natural Language Processing and the 5th Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics (Volume 1: Long Papers)
%D 2026
%8 November
%I Association for Computational Linguistics
%C Hengqin, China
%@ 979-8-89176-439-2
%F shibata-miyamura-2026-preferences
%X Multi-trait automated essay scoring (AES) provides trait-level scores as feedback but requires costly annotations across multiple traits to train accurate scoring models. To alleviate this annotation burden, we propose MTPAS (Multi-Trait Preference-to-Absolute Scoring), a semi-supervised framework that learns relative score orderings from LLM-generated pairwise preferences and calibrates them to absolute rubric scales using a small set of human-scored anchor essays in an end-to-end manner. To this end, MTPAS jointly optimizes three objectives: a Bradley–Terry pairwise preference loss, a regression loss on the anchor essays, and a Sinkhorn divergence-based distribution regularizer. The model adopts a bifactor architecture that decomposes essay representations into a shared general factor and trait-specific residuals, improving robustness against noisy LLM-generated preferences. We further introduce pairwise rubric compression, which reformulates the original score-level rubrics into concise comparative criteria, reducing the inference cost of collecting LLM preferences. On ASAP and ASAP++, MTPAS achieves 91.4% of the fully supervised SoTA performance with only 30 anchor essays per prompt and further narrows the gap with 200 anchors. Under the same anchor budget, MTPAS outperforms conventional multi-trait AES baselines at every anchor size investigated in this study.
%U https://aclanthology.org/2026.aacl-long.273/
%P 6041-6057

Markdown (Informal)

[From Preferences to Scores: Semi-Supervised Multi-Trait Essay Scoring with LLM Pairwise Preferences and Few Human Anchors](https://aclanthology.org/2026.aacl-long.273/) (Shibata & Miyamura, AACL-IJCNLP 2026)

• From Preferences to Scores: Semi-Supervised Multi-Trait Essay Scoring with LLM Pairwise Preferences and Few Human Anchors (Shibata & Miyamura, AACL-IJCNLP 2026)

ACL

• Takumi Shibata and Yuichi Miyamura. 2026. From Preferences to Scores: Semi-Supervised Multi-Trait Essay Scoring with LLM Pairwise Preferences and Few Human Anchors. In Proceedings of the 15th International Joint Conference on Natural Language Processing and the 5th Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6041–6057, Hengqin, China. Association for Computational Linguistics.