@inproceedings{kurfali-2026-swespecqa,
title = "{S}we{S}pec{QA}: A {S}wedish Benchmark for Specialized Document-Native Question Answering",
author = "Kurfali, Murathan",
editor = "Zhang, Yue and
Xu, Ruifeng and
Purwarianti, Ayu and
Rong, Tong and
Suzuki, Jun",
booktitle = "Proceedings of the 15th International Joint Conference on Natural Language Processing and the 5th Conference of the {A}sia-Pacific Chapter of the {A}ssociation for {C}omputational {L}inguistics (Volume 1: Long Papers)",
month = nov,
year = "2026",
address = "Hengqin, China",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.aacl-long.92/",
pages = "1998--2022",
ISBN = "979-8-89176-439-2",
abstract = "We introduce \textit{SweSpecQA}, to our knowledge, the first Swedish benchmark for specialized institutional and professional QA with paired text and full-page PDF evaluation. It contains 1,500 questions from public Swedish examination and certification materials across six domains: EU aptitude, boating, antibiotic stewardship, interpreter licensing, physician licensing, and pharmacist licensing. Of these, 234 questions require an original visual or layout-dependent source element; the remaining 1,266 are text-eligible. All benchmark questions are drawn directly from Swedish-language source materials, avoiding translation artifacts in multilingual adaptations of English benchmarks. We evaluate 17 models under a closed-book protocol, reporting accuracy by domain, exam subtype, visual/layout dependence, modality, and prompting strategy, as well as whether models follow the requested answer format. On the text-eligible subset, every multimodal model performs worse with full source-page input than extracted text; PDF penalties range from 4.0 to 38.2 points under direct prompting and from 1.7 to 36.4 points under chain-of-thought prompting. Reasoning prompts often improve PDF performance for several open-weight VLMs, while sometimes reducing clean-text accuracy. These results show that document-native evaluation exposes failures hidden by text-only evaluation and that preprocessing remains important for reliable Swedish exam-style QA."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="kurfali-2026-swespecqa">
<titleInfo>
<title>SweSpecQA: A Swedish Benchmark for Specialized Document-Native Question Answering</title>
</titleInfo>
<name type="personal">
<namePart type="given">Murathan</namePart>
<namePart type="family">Kurfali</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>We introduce SweSpecQA, to our knowledge, the first Swedish benchmark for specialized institutional and professional QA with paired text and full-page PDF evaluation. It contains 1,500 questions from public Swedish examination and certification materials across six domains: EU aptitude, boating, antibiotic stewardship, interpreter licensing, physician licensing, and pharmacist licensing. Of these, 234 questions require an original visual or layout-dependent source element; the remaining 1,266 are text-eligible. All benchmark questions are drawn directly from Swedish-language source materials, avoiding translation artifacts in multilingual adaptations of English benchmarks. We evaluate 17 models under a closed-book protocol, reporting accuracy by domain, exam subtype, visual/layout dependence, modality, and prompting strategy, as well as whether models follow the requested answer format. On the text-eligible subset, every multimodal model performs worse with full source-page input than extracted text; PDF penalties range from 4.0 to 38.2 points under direct prompting and from 1.7 to 36.4 points under chain-of-thought prompting. Reasoning prompts often improve PDF performance for several open-weight VLMs, while sometimes reducing clean-text accuracy. These results show that document-native evaluation exposes failures hidden by text-only evaluation and that preprocessing remains important for reliable Swedish exam-style QA.</abstract>
<identifier type="citekey">kurfali-2026-swespecqa</identifier>
<location>
<url>https://aclanthology.org/2026.aacl-long.92/</url>
</location>
<part>
<date>2026-11</date>
<extent unit="page">
<start>1998</start>
<end>2022</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T SweSpecQA: A Swedish Benchmark for Specialized Document-Native Question Answering
%A Kurfali, Murathan
%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 kurfali-2026-swespecqa
%X We introduce SweSpecQA, to our knowledge, the first Swedish benchmark for specialized institutional and professional QA with paired text and full-page PDF evaluation. It contains 1,500 questions from public Swedish examination and certification materials across six domains: EU aptitude, boating, antibiotic stewardship, interpreter licensing, physician licensing, and pharmacist licensing. Of these, 234 questions require an original visual or layout-dependent source element; the remaining 1,266 are text-eligible. All benchmark questions are drawn directly from Swedish-language source materials, avoiding translation artifacts in multilingual adaptations of English benchmarks. We evaluate 17 models under a closed-book protocol, reporting accuracy by domain, exam subtype, visual/layout dependence, modality, and prompting strategy, as well as whether models follow the requested answer format. On the text-eligible subset, every multimodal model performs worse with full source-page input than extracted text; PDF penalties range from 4.0 to 38.2 points under direct prompting and from 1.7 to 36.4 points under chain-of-thought prompting. Reasoning prompts often improve PDF performance for several open-weight VLMs, while sometimes reducing clean-text accuracy. These results show that document-native evaluation exposes failures hidden by text-only evaluation and that preprocessing remains important for reliable Swedish exam-style QA.
%U https://aclanthology.org/2026.aacl-long.92/
%P 1998-2022
Markdown (Informal)
[SweSpecQA: A Swedish Benchmark for Specialized Document-Native Question Answering](https://aclanthology.org/2026.aacl-long.92/) (Kurfali, AACL-IJCNLP 2026)
ACL