<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="zahid-etal-2026-modality">
    <titleInfo>
        <title>Modality-Conditioned Ensembles and Zero-Shot Cross-Encoder Reranking for Relation Classification Between Astronomy-Paper Objects</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">M</namePart>
        <namePart type="given">Tahsan</namePart>
        <namePart type="family">Zahid</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Most.</namePart>
        <namePart type="given">Atkia</namePart>
        <namePart type="family">Farzana</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Tonmoy</namePart>
        <namePart type="given">Sarker</namePart>
        <namePart type="family">Sourav</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 4th Workshop on Artificial Intelligence for Scientific Publications</title>
        </titleInfo>
        <name type="personal">
            <namePart type="given">Alberto</namePart>
            <namePart type="family">Accomazzi</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Tirthankar</namePart>
            <namePart type="family">Ghosal</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Felix</namePart>
            <namePart type="family">Grezes</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Kelly</namePart>
            <namePart type="family">Lockhart</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Atilla</namePart>
            <namePart type="given">Kaan</namePart>
            <namePart type="family">Alkan</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <originInfo>
            <publisher>Association for Computational Linguistics</publisher>
            <place>
                <placeTerm type="text">Hengqin, Zhuhai, China</placeTerm>
            </place>
        </originInfo>
        <genre authority="marcgt">conference publication</genre>
        <identifier type="isbn">979-8-89176-456-9</identifier>
    </relatedItem>
    <abstract>The WASP 2026 (AstroCLIMB) task requires classifying the relation between a pair of objects drawn from astronomy papers — same_figure, same_paper, related_papers, or unrelated_papers — where each pair combines text (captions) and/or images (figures) in one of three disjoint modality configurations: image+text, text+text, and image+image. Since same_figure cannot occur outside the image+text subset, a single unified classifier is structurally sub-optimal; we instead adopt three independent per-modality classifiers whose predictions are pooled for the competition metric (pooled macro-F1). Starting from a hand-engineered baseline (CLIP, SciNCL and DINOv2 embeddings with lexical/OCR features and a LightGBM ensemble), we find that replacing the encoder with a joint vision-language embedding model (Qwen3-VL-Embedding-2B) produces the largest single improvement among the experiments we tested (public leaderboard 0.4808 → 0.62319), exceeding the combined effect of six feature-engineering ideas that were implemented and measured but produced less than 0.01 macro-F1 each. We further show that zero-shot cross-encoder rerankers (Qwen3-VL-Reranker-2B, Qwen3-Reranker-4B), applied with no task-specific fine-tuning, add a second large improvement on top of the engineered ensemble (leaderboard 0.64210 → 0.66299), and that an untrained text+text reranker alone already outperforms the fully engineered pipeline on that subset (macro-F1 0.7111 vs. 0.6704). We report a parallel distant-supervision branch that trains on 120,000 externally-labeled pairs mined from the competition’s source corpus and reaches a comparable but lower leaderboard score (0.57713) through a different mechanism; under our tested data and modeling regimes, this suggests that representation quality was a stronger bottleneck than either additional hand-engineered features or additional distant supervision. All reported numbers are traced to raw out-of-fold prediction arrays or literal notebook output under a shared leak-free StratifiedKFold(5) protocol, and we document two negative results (a broken nested-honest validation split, and several feature-engineering ideas that measurably hurt) as concrete engineering findings.</abstract>
    <identifier type="citekey">zahid-etal-2026-modality</identifier>
    <location>
        <url>https://aclanthology.org/2026.wasp-1.13/</url>
    </location>
    <part>
        <date>2026-11</date>
        <extent unit="page">
            <start>112</start>
            <end>121</end>
        </extent>
    </part>
</mods>
</modsCollection>

%0 Conference Proceedings
%T Modality-Conditioned Ensembles and Zero-Shot Cross-Encoder Reranking for Relation Classification Between Astronomy-Paper Objects
%A Zahid, M. Tahsan
%A Farzana, Most. Atkia
%A Sourav, Tonmoy Sarker
%Y Accomazzi, Alberto
%Y Ghosal, Tirthankar
%Y Grezes, Felix
%Y Lockhart, Kelly
%Y Alkan, Atilla Kaan
%S Proceedings of the 4th Workshop on Artificial Intelligence for Scientific Publications
%D 2026
%8 November
%I Association for Computational Linguistics
%C Hengqin, Zhuhai, China
%@ 979-8-89176-456-9
%F zahid-etal-2026-modality
%X The WASP 2026 (AstroCLIMB) task requires classifying the relation between a pair of objects drawn from astronomy papers — same_figure, same_paper, related_papers, or unrelated_papers — where each pair combines text (captions) and/or images (figures) in one of three disjoint modality configurations: image+text, text+text, and image+image. Since same_figure cannot occur outside the image+text subset, a single unified classifier is structurally sub-optimal; we instead adopt three independent per-modality classifiers whose predictions are pooled for the competition metric (pooled macro-F1). Starting from a hand-engineered baseline (CLIP, SciNCL and DINOv2 embeddings with lexical/OCR features and a LightGBM ensemble), we find that replacing the encoder with a joint vision-language embedding model (Qwen3-VL-Embedding-2B) produces the largest single improvement among the experiments we tested (public leaderboard 0.4808 → 0.62319), exceeding the combined effect of six feature-engineering ideas that were implemented and measured but produced less than 0.01 macro-F1 each. We further show that zero-shot cross-encoder rerankers (Qwen3-VL-Reranker-2B, Qwen3-Reranker-4B), applied with no task-specific fine-tuning, add a second large improvement on top of the engineered ensemble (leaderboard 0.64210 → 0.66299), and that an untrained text+text reranker alone already outperforms the fully engineered pipeline on that subset (macro-F1 0.7111 vs. 0.6704). We report a parallel distant-supervision branch that trains on 120,000 externally-labeled pairs mined from the competition’s source corpus and reaches a comparable but lower leaderboard score (0.57713) through a different mechanism; under our tested data and modeling regimes, this suggests that representation quality was a stronger bottleneck than either additional hand-engineered features or additional distant supervision. All reported numbers are traced to raw out-of-fold prediction arrays or literal notebook output under a shared leak-free StratifiedKFold(5) protocol, and we document two negative results (a broken nested-honest validation split, and several feature-engineering ideas that measurably hurt) as concrete engineering findings.
%U https://aclanthology.org/2026.wasp-1.13/
%P 112-121

Markdown (Informal)

[Modality-Conditioned Ensembles and Zero-Shot Cross-Encoder Reranking for Relation Classification Between Astronomy-Paper Objects](https://aclanthology.org/2026.wasp-1.13/) (Zahid et al., WASP 2026)

• Modality-Conditioned Ensembles and Zero-Shot Cross-Encoder Reranking for Relation Classification Between Astronomy-Paper Objects (Zahid et al., WASP 2026)

ACL

• M. Tahsan Zahid, Most. Atkia Farzana, and Tonmoy Sarker Sourav. 2026. Modality-Conditioned Ensembles and Zero-Shot Cross-Encoder Reranking for Relation Classification Between Astronomy-Paper Objects. In Proceedings of the 4th Workshop on Artificial Intelligence for Scientific Publications, pages 112–121, Hengqin, Zhuhai, China. Association for Computational Linguistics.