Manabu Okumura
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Manabu Okumura has a total of 102 co-authors. Here are the top co-authors by number of shared publications:
- Hiroya Takamura
National Institute of Advanced Industrial Science and Technology
10 shared publications - Dongyuan Li
Center for Spatial Information Science, The University of Tokyo, Tokyo, Japan
8 shared publications - Hidetaka Kamigaito
Nara Institute of Science and Technology (NAIST) Japan
7 shared publications - Renhe Jiang
The University of Tokyo, Tokyo, Japan
6 shared publications - Jingun Kwon
Chungnam National University
5 shared publications - Zhen Wang
School of Information and Communication Engineering, Tokyo Institute of Technology, Tokyo, Japan
5 shared publications
Latest Publications
- Open Access
- 9 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 10335-10344
Simile Understanding in Text-to-Image Models: An Evaluation Framework
- 9 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 10335-10344
Open AccessSimile Understanding in Text-to-Image Models: An Evaluation Framework
- Open Access
- 26 May 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 94377-94392
CodeNER: Code Prompting for Named Entity Recognition
Recent studies have explored various approaches for treating candidate named entity spans as both source and target sequences in named entity recognition (NER) by leveraging large language models (LLMs). Although previous approaches have successfully generated candidate named entity spans with suitable labels, they rely solely on the input context information when using LLMs, particularly, ChatGPT. However, NER inherently requires capturing detailed labeling requirements with input context information. To address this issue, we propose a novel method that leverages code-based prompting to improve the capabilities of LLMs in understanding and performing NER. By embedding code within prompts, we provide detailed Beginning-Inside-Outside (BIO) schema instructions for labeling, thereby exploiting the ability of LLMs to comprehend long-range scopes in programming languages. Experimental results demonstrate that the proposed code-based prompting method outperforms conventional text-based prompting on ten benchmarks across English, Arabic, Finnish, Danish, and German datasets, indicating the effectiveness of explicitly structuring NER instructions. Furthermore, our in-depth analysis showed that the performance improvement of the proposed method stems from code-based structuring and BIO-guided processing rather than from simple entity-label descriptions. Moreover, the method maintained its effectiveness or demonstrated additional performance improvements when combined with Chain-of-Thought (CoT) prompting and few-shot settings.
- 26 May 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 94377-94392
Open AccessCodeNER: Code Prompting for Named Entity Recognition
Recent studies have explored various approaches for treating candidate named entity spans as both source and target sequences in named entity recognition (NER) by leveraging large language models (LLMs). Although previous approaches have successfully generated candidate named entity spans with suitable labels, they rely solely on the input context information when using LLMs, particularly, ChatGPT. However, NER inherently requires capturing detailed labeling requirements with input context information. To address this issue, we propose a novel method that leverages code-based prompting to improve the capabilities of LLMs in understanding and performing NER. By embedding code within prompts, we provide detailed Beginning-Inside-Outside (BIO) schema instructions for labeling, thereby exploiting the ability of LLMs to comprehend long-range scopes in programming languages. Experimental results demonstrate that the proposed code-based prompting method outperforms conventional text-based prompting on ten benchmarks across English, Arabic, Finnish, Danish, and German datasets, indicating the effectiveness of explicitly structuring NER instructions. Furthermore, our in-depth analysis showed that the performance improvement of the proposed method stems from code-based structuring and BIO-guided processing rather than from simple entity-label descriptions. Moreover, the method maintained its effectiveness or demonstrated additional performance improvements when combined with Chain-of-Thought (CoT) prompting and few-shot settings.
- Open Access
- 11 May 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 72226-72243
Biclique-Guided Graph Contrastive Learning for Recommendation
Graph Contrastive Learning (GCL) has been widely adopted in recommendation systems to learn user and item representations from implicit user–item interactions. However, standard GCL formulations define positive pairs as the same node across augmented views and treat all other nodes as negatives, leading to critical limitations in user–item interaction graphs. Specifically, such designs introduce false negatives by contrasting users or items with highly similar interaction patterns and lack explicit semantic guidance, since contrastive pairs are determined implicitly by augmentation mechanics rather than meaningful interaction structures. To address these issues, we propose a biclique-guided graph contrastive learning framework that explicitly leverages user–item interaction structures to guide positive and negative sample selection. We model groups of users and items as soft bicliques, which act as relaxed structural prototypes capturing consistent collective preference patterns. By relaxing the strict connectivity constraints of hard bicliques, our formulation avoids NP-hard exact biclique discovery and remains effective under sparse interactions and missing edges. Nodes within the same soft biclique form positive contrastive pairs, while nodes across different bicliques provide more reliable negatives, injecting explicit semantic supervision beyond instance-level or clustering-based heuristics. Extensive experiments on multiple benchmark datasets demonstrate consistent performance improvements and superior representation quality over state-of-the-art contrastive recommendation methods.
- 11 May 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 72226-72243
Open AccessBiclique-Guided Graph Contrastive Learning for Recommendation
Graph Contrastive Learning (GCL) has been widely adopted in recommendation systems to learn user and item representations from implicit user–item interactions. However, standard GCL formulations define positive pairs as the same node across augmented views and treat all other nodes as negatives, leading to critical limitations in user–item interaction graphs. Specifically, such designs introduce false negatives by contrasting users or items with highly similar interaction patterns and lack explicit semantic guidance, since contrastive pairs are determined implicitly by augmentation mechanics rather than meaningful interaction structures. To address these issues, we propose a biclique-guided graph contrastive learning framework that explicitly leverages user–item interaction structures to guide positive and negative sample selection. We model groups of users and items as soft bicliques, which act as relaxed structural prototypes capturing consistent collective preference patterns. By relaxing the strict connectivity constraints of hard bicliques, our formulation avoids NP-hard exact biclique discovery and remains effective under sparse interactions and missing edges. Nodes within the same soft biclique form positive contrastive pairs, while nodes across different bicliques provide more reliable negatives, injecting explicit semantic supervision beyond instance-level or clustering-based heuristics. Extensive experiments on multiple benchmark datasets demonstrate consistent performance improvements and superior representation quality over state-of-the-art contrastive recommendation methods.
- Open Access
MMCIG: Multimodal Cover Image Generation for Text-only Documents and Its Dataset Construction via Pseudo-labeling
In this study, we introduce a novel cover image generation task that produces both a concise summary and a visually corresponding image from a given text-only document. Because no existing datasets are available for this task, we propose a multimodal pseudo-labeling method to construct high-quality datasets at low cost. We first collect documents that contain multiple images with their captions, and their summaries by excluding factually inconsistent instances. Our approach selects one image from the multiple images accompanying the documents. Using the gold summary, we independently rank both the images and their captions. Then, we annotate a pseudo-label for an image when both the image and its corresponding caption are ranked first in their respective rankings. Finally, we remove documents that contain direct image references within texts. Experimental results demonstrate that the proposed multimodal pseudo-labeling method constructs more precise datasets and generates higher quality images than text- and image-only pseudo-labeling methods, which consider captions and images separately. We release our code at: https://github.com/HyeyeeonKim/MMCIG
Open AccessMMCIG: Multimodal Cover Image Generation for Text-only Documents and Its Dataset Construction via Pseudo-labeling
In this study, we introduce a novel cover image generation task that produces both a concise summary and a visually corresponding image from a given text-only document. Because no existing datasets are available for this task, we propose a multimodal pseudo-labeling method to construct high-quality datasets at low cost. We first collect documents that contain multiple images with their captions, and their summaries by excluding factually inconsistent instances. Our approach selects one image from the multiple images accompanying the documents. Using the gold summary, we independently rank both the images and their captions. Then, we annotate a pseudo-label for an image when both the image and its corresponding caption are ranked first in their respective rankings. Finally, we remove documents that contain direct image references within texts. Experimental results demonstrate that the proposed multimodal pseudo-labeling method constructs more precise datasets and generates higher quality images than text- and image-only pseudo-labeling methods, which consider captions and images separately. We release our code at: https://github.com/HyeyeeonKim/MMCIG
- Open Access
- 3 August 2025
- Published by Association for Computing Machinery (ACM)
- p. 2790-2800
Taming Recommendation Bias with Causal Intervention on Evolving Personal Popularity
- 3 August 2025
- Published by Association for Computing Machinery (ACM)
- p. 2790-2800
Open AccessTaming Recommendation Bias with Causal Intervention on Evolving Personal Popularity
- Open Access
- 13 July 2025
- Published by Association for Computing Machinery (ACM)
- p. 1423-1433
A Unified Retrieval Framework with Document Ranking and EDU Filtering for Multi-document Summarization
- 13 July 2025
- Published by Association for Computing Machinery (ACM)
- p. 1423-1433
Open AccessA Unified Retrieval Framework with Document Ranking and EDU Filtering for Multi-document Summarization
- Open Access
- 22 April 2025
- Published by Association for Computing Machinery (ACM)
- p. 1342-1352
Revisiting Dynamic Graph Clustering via Matrix Factorization
- 22 April 2025
- Published by Association for Computing Machinery (ACM)
- p. 1342-1352
Open AccessRevisiting Dynamic Graph Clustering via Matrix Factorization
- Open Access
Considering Length Diversity in Retrieval-Augmented Summarization
This study investigates retrieval-augmented summarization by specifically examining the impact of exemplar summary lengths under length constraints, not covered by previous work. We propose a Diverse Length-aware Maximal Marginal Relevance (DL-MMR) algorithm to better control summary lengths. This algorithm combines the query relevance with diverse target lengths in retrieval-augmented summarization. Unlike previous methods that necessitate exhaustive exemplar exemplar relevance comparisons using MMR, DL-MMR considers the exemplar target length as well and avoids comparing exemplars to each other, thereby reducing computational cost and conserving memory during the construction of an exemplar pool. Experimental results showed the effectiveness of DL-MMR, which considers length diversity, compared to the original MMR algorithm. DL-MMR additionally showed the effectiveness in memory saving of 781,513 times and computational cost reduction of 500,092 times, while maintaining the same level of informativeness.
Open AccessConsidering Length Diversity in Retrieval-Augmented Summarization
This study investigates retrieval-augmented summarization by specifically examining the impact of exemplar summary lengths under length constraints, not covered by previous work. We propose a Diverse Length-aware Maximal Marginal Relevance (DL-MMR) algorithm to better control summary lengths. This algorithm combines the query relevance with diverse target lengths in retrieval-augmented summarization. Unlike previous methods that necessitate exhaustive exemplar exemplar relevance comparisons using MMR, DL-MMR considers the exemplar target length as well and avoids comparing exemplars to each other, thereby reducing computational cost and conserving memory during the construction of an exemplar pool. Experimental results showed the effectiveness of DL-MMR, which considers length diversity, compared to the original MMR algorithm. DL-MMR additionally showed the effectiveness in memory saving of 781,513 times and computational cost reduction of 500,092 times, while maintaining the same level of informativeness.
- Top Cited PapersOpen Access
- 6 January 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 13, 5491-5506
Continuous Sign Language Recognition With Multi-Scale Spatial-Temporal Feature Enhancement
Continuous Sign Language Recognition (CSLR) seeks to interpret the gestures used by deaf-mute individuals and translate them into natural language, thereby enhancing communication and interaction. A successful CSLR method relies on the continuous tracking of the presenter’s gestures and facial movements. Existing CSLR methods struggle with fully leveraging fine-grained continuous frame information and often overlook the importance of multi-scale feature integration during decoding. To solve the above-mentioned issues, in this paper, we propose a spatial-temporal feature-enhanced network, called STNet for CSLR task. Firstly, for better continuous frame information exploration, based on the optimal transport algorithm, we first propose a spatial resonance module, which is used to extract the global common spatial features of two adjacent frames along the frame sequence. Secondly, we design a frame-wise loss to preserve and enhance the specific features of each frame. Lastly, to emphasize the multi-scale feature fusion, on the decoder side, we design a multi-temporal perception module, to allow each frame to focus on a larger range of other frames and enhance information interaction from different scales. Extensive experiments on three benchmark datasets including PHOENIX14, PHOENIX14-T, and CSL-Daily demonstrate that STNet consistently outperforms state-of-the-art methods, with a notable improvement of 2.9% in CSLR, showcasing its effectiveness and generalizability. Our approach provides a robust foundation for real-world applications such as sign language education and communication tools, while ablation and case studies highlight the impact of each module, paving the way for future research in CSLR.
- 6 January 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 13, 5491-5506
Top Cited PapersOpen AccessContinuous Sign Language Recognition With Multi-Scale Spatial-Temporal Feature Enhancement
Continuous Sign Language Recognition (CSLR) seeks to interpret the gestures used by deaf-mute individuals and translate them into natural language, thereby enhancing communication and interaction. A successful CSLR method relies on the continuous tracking of the presenter’s gestures and facial movements. Existing CSLR methods struggle with fully leveraging fine-grained continuous frame information and often overlook the importance of multi-scale feature integration during decoding. To solve the above-mentioned issues, in this paper, we propose a spatial-temporal feature-enhanced network, called STNet for CSLR task. Firstly, for better continuous frame information exploration, based on the optimal transport algorithm, we first propose a spatial resonance module, which is used to extract the global common spatial features of two adjacent frames along the frame sequence. Secondly, we design a frame-wise loss to preserve and enhance the specific features of each frame. Lastly, to emphasize the multi-scale feature fusion, on the decoder side, we design a multi-temporal perception module, to allow each frame to focus on a larger range of other frames and enhance information interaction from different scales. Extensive experiments on three benchmark datasets including PHOENIX14, PHOENIX14-T, and CSL-Daily demonstrate that STNet consistently outperforms state-of-the-art methods, with a notable improvement of 2.9% in CSLR, showcasing its effectiveness and generalizability. Our approach provides a robust foundation for real-world applications such as sign language education and communication tools, while ablation and case studies highlight the impact of each module, paving the way for future research in CSLR.
- Open Access
- 1 January 2025
- Published by Association for Computational Linguistics (ACL)
- p. 19775-19793
Length Representations in Large Language Models
Sangjun Moon, Dasom Choi, Jingun Kwon, Hidetaka Kamigaito, Manabu Okumura. Findings of the Association for Computational Linguistics: EMNLP 2025. 2025.
- 1 January 2025
- Published by Association for Computational Linguistics (ACL)
- p. 19775-19793
Open AccessLength Representations in Large Language Models
Sangjun Moon, Dasom Choi, Jingun Kwon, Hidetaka Kamigaito, Manabu Okumura. Findings of the Association for Computational Linguistics: EMNLP 2025. 2025.