@inproceedings{katsube-etal-2026-dejima,
title = "{DEJIMA}: A Novel Large-scale {J}apanese Dataset for Image Captioning and Visual Question Answering",
author = "Katsube, Toshiki and
Taiga, Fukuhara and
Ando, Kenichiro and
Mukuta, Yusuke and
Uehara, Kohei and
Harada, Tatsuya",
editor = "Piperidis, Stelios and
Bel, N{\'u}ria and
van den Heuvel, Henk and
Ide, Nancy and
Krek, Simon and
Toral, Antonio",
booktitle = "Proceedings of the Fifteenth Language Resources and Evaluation Conference",
month = may,
year = "2026",
address = "Palma de Mallorca, Spain",
publisher = "ELRA Language Resource Association",
url = "https://aclanthology.org/2026.lrec-1.744/",
doi = "10.63317/45nioi7qjz28",
pages = "9478--9489",
abstract = "Vision-and-Language (V{\&}L) models depend on large-scale, high-quality datasets, yet most resources are English-centric, and existing Japanese V{\&}L datasets face a fundamental trade-off: manually annotated corpora offer quality but limited scale, translated datasets introduce unnatural phrasing and cultural bias, and web-crawled collections achieve scale but suffer from noise and poor grounding. To resolve this trade-off, we propose DEJIMA, a novel pipeline whose key idea is detection-guided LLM refinement: object detection first extracts visually verifiable evidence (labels and bounding boxes), then an LLM generates or refines Japanese text conditioned on this evidence, ensuring both factual grounding and linguistic naturalness without costly human annotation. Using this pipeline, we build two resources: an image{--}caption dataset (DEJIMA-Cap) and a VQA dataset (DEJIMA-VQA), each containing approximately 3.88M image{--}text pairs{---}over 20 times larger than existing Japanese V{\&}L datasets. Human evaluations demonstrate that DEJIMA achieves substantially higher Japaneseness and linguistic naturalness than translation- or annotation-based baselines, while maintaining factual correctness comparable to human-annotated corpora. Models trained on DEJIMA show consistent improvements across multiple Japanese multimodal benchmarks, confirming that culturally grounded, large-scale resources play a key role in enhancing model performance. All pipeline components are commercially licensed, and we publicly release the dataset and metadata to support further research and applications. Our project page is available at \url{https://mil-tokyo.github.io/DEJIMA-dataset/}."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="katsube-etal-2026-dejima">
<titleInfo>
<title>DEJIMA: A Novel Large-scale Japanese Dataset for Image Captioning and Visual Question Answering</title>
</titleInfo>
<name type="personal">
<namePart type="given">Toshiki</namePart>
<namePart type="family">Katsube</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Fukuhara</namePart>
<namePart type="family">Taiga</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Kenichiro</namePart>
<namePart type="family">Ando</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yusuke</namePart>
<namePart type="family">Mukuta</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Kohei</namePart>
<namePart type="family">Uehara</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tatsuya</namePart>
<namePart type="family">Harada</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-05</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the Fifteenth Language Resources and Evaluation Conference</title>
</titleInfo>
<name type="personal">
<namePart type="given">Stelios</namePart>
<namePart type="family">Piperidis</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Núria</namePart>
<namePart type="family">Bel</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Henk</namePart>
<namePart type="family">van den Heuvel</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Nancy</namePart>
<namePart type="family">Ide</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Simon</namePart>
<namePart type="family">Krek</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Antonio</namePart>
<namePart type="family">Toral</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>ELRA Language Resource Association</publisher>
<place>
<placeTerm type="text">Palma de Mallorca, Spain</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
</relatedItem>
<abstract>Vision-and-Language (V&L) models depend on large-scale, high-quality datasets, yet most resources are English-centric, and existing Japanese V&L datasets face a fundamental trade-off: manually annotated corpora offer quality but limited scale, translated datasets introduce unnatural phrasing and cultural bias, and web-crawled collections achieve scale but suffer from noise and poor grounding. To resolve this trade-off, we propose DEJIMA, a novel pipeline whose key idea is detection-guided LLM refinement: object detection first extracts visually verifiable evidence (labels and bounding boxes), then an LLM generates or refines Japanese text conditioned on this evidence, ensuring both factual grounding and linguistic naturalness without costly human annotation. Using this pipeline, we build two resources: an image–caption dataset (DEJIMA-Cap) and a VQA dataset (DEJIMA-VQA), each containing approximately 3.88M image–text pairs—over 20 times larger than existing Japanese V&L datasets. Human evaluations demonstrate that DEJIMA achieves substantially higher Japaneseness and linguistic naturalness than translation- or annotation-based baselines, while maintaining factual correctness comparable to human-annotated corpora. Models trained on DEJIMA show consistent improvements across multiple Japanese multimodal benchmarks, confirming that culturally grounded, large-scale resources play a key role in enhancing model performance. All pipeline components are commercially licensed, and we publicly release the dataset and metadata to support further research and applications. Our project page is available at https://mil-tokyo.github.io/DEJIMA-dataset/.</abstract>
<identifier type="citekey">katsube-etal-2026-dejima</identifier>
<identifier type="doi">10.63317/45nioi7qjz28</identifier>
<location>
<url>https://aclanthology.org/2026.lrec-1.744/</url>
</location>
<part>
<date>2026-05</date>
<extent unit="page">
<start>9478</start>
<end>9489</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T DEJIMA: A Novel Large-scale Japanese Dataset for Image Captioning and Visual Question Answering
%A Katsube, Toshiki
%A Taiga, Fukuhara
%A Ando, Kenichiro
%A Mukuta, Yusuke
%A Uehara, Kohei
%A Harada, Tatsuya
%Y Piperidis, Stelios
%Y Bel, Núria
%Y van den Heuvel, Henk
%Y Ide, Nancy
%Y Krek, Simon
%Y Toral, Antonio
%S Proceedings of the Fifteenth Language Resources and Evaluation Conference
%D 2026
%8 May
%I ELRA Language Resource Association
%C Palma de Mallorca, Spain
%F katsube-etal-2026-dejima
%X Vision-and-Language (V&L) models depend on large-scale, high-quality datasets, yet most resources are English-centric, and existing Japanese V&L datasets face a fundamental trade-off: manually annotated corpora offer quality but limited scale, translated datasets introduce unnatural phrasing and cultural bias, and web-crawled collections achieve scale but suffer from noise and poor grounding. To resolve this trade-off, we propose DEJIMA, a novel pipeline whose key idea is detection-guided LLM refinement: object detection first extracts visually verifiable evidence (labels and bounding boxes), then an LLM generates or refines Japanese text conditioned on this evidence, ensuring both factual grounding and linguistic naturalness without costly human annotation. Using this pipeline, we build two resources: an image–caption dataset (DEJIMA-Cap) and a VQA dataset (DEJIMA-VQA), each containing approximately 3.88M image–text pairs—over 20 times larger than existing Japanese V&L datasets. Human evaluations demonstrate that DEJIMA achieves substantially higher Japaneseness and linguistic naturalness than translation- or annotation-based baselines, while maintaining factual correctness comparable to human-annotated corpora. Models trained on DEJIMA show consistent improvements across multiple Japanese multimodal benchmarks, confirming that culturally grounded, large-scale resources play a key role in enhancing model performance. All pipeline components are commercially licensed, and we publicly release the dataset and metadata to support further research and applications. Our project page is available at https://mil-tokyo.github.io/DEJIMA-dataset/.
%R 10.63317/45nioi7qjz28
%U https://aclanthology.org/2026.lrec-1.744/
%U https://doi.org/10.63317/45nioi7qjz28
%P 9478-9489
Markdown (Informal)
[DEJIMA: A Novel Large-scale Japanese Dataset for Image Captioning and Visual Question Answering](https://aclanthology.org/2026.lrec-1.744/) (Katsube et al., LREC 2026)
ACL