Papers by Shuhei Kurita

13 papers
Demystifying Mixed Outcomes of Self-Training: Pre-training Analyses on Non-Toy LLMs (2026.findings-eacl)

Copied to clipboard

Challenge: Recent studies on self-training report seemingly contradictory outcomes.
Approach: They use OLMo-2 models as non-toy LLMs and perform multiple rounds of continual pre-training using self-generated text with different prompting strategies and data filtering.
Outcome: The proposed model collapse is inherent to the training procedure itself, while self-improvement is likely owes its success to human-designed, strategic synthetic pipelines that inject external intelligence.
Co-Teaching Student-Model through Submission Results of Shared Task (2021.findings-emnlp)

Copied to clipboard

Challenge: Shared tasks require participants to submit only system outputs and descriptions.
Approach: They propose to utilize all system outputs in a shared task to build a unified system that performs better than the task's single best system.
Outcome: The proposed scheme outperforms the best system in the SHINRA2019-JP shared task with nine participants.
Developing Japanese CLIP Models Leveraging an Open-weight LLM for Large-scale Dataset Translation (2025.naacl-srw)

Copied to clipboard

Challenge: lack of large-scale open Japanese image-text pairs poses a significant barrier to the development of vision-language models.
Approach: They construct large-scale Japanese image-text pairs using machine translation and pre-trained CLIP models on a Japanese dataset.
Outcome: The results show that pre-trained models achieve competitive average scores on Japanese culture tasks compared to models of similar size.
Visual Recipe Flow: A Dataset for Learning Visual State Changes of Objects with Recipe Flows (2022.coling-1)

Copied to clipboard

Challenge: a new dataset enables us to learn a cooking action result for each object in a recipe text.
Approach: They propose a multimodal dataset that enables us to learn a cooking action result for each object in a recipe text.
Outcome: The proposed dataset reduces human annotation costs by allowing multimodal information retrieval.
Query-based Image Captioning from Multi-context 360cdegree Images (2023.findings-emnlp)

Copied to clipboard

Challenge: Existing image captioning datasets follow salient images with limited camera field of view, ignoring minor details.
Approach: They propose a task where a query specifies the context to describe in 360-degree images and construct a dataset for the task that contains 3,940 360- degree images and 18,459 pairs of queries and captions annotated manually.
Outcome: The proposed task is more challenging than the conventional image captioning task, which describes salient objects in images.
LegalViz: Legal Text Visualization by Text To Diagram Generation (2025.naacl-long)

Copied to clipboard

Challenge: Graphviz provides diagrams for legal documents that are easy to understand and understand . a novel dataset of 23 languages and 7,010 cases of legal document and visualization pairs is proposed .
Approach: They propose a dataset of legal diagrams using DOT graph description language of Graphviz.
Outcome: The proposed dataset outperforms existing models including GPTs in 23 languages and 7,010 cases of legal document and visualization pairs.
Text360Nav: 360-Degree Image Captioning Dataset for Urban Pedestrians Navigation (2024.lrec-main)

Copied to clipboard

Challenge: Existing image captioning datasets focus on the overall image description and lack detailed scene descriptions, overlooking features for pedestrians walking on urban streets.
Approach: They develop a dataset to provide textual feedback from 360-degree camera images to visually impaired pedestrians . they generate meaningful captions focusing on obstacles on the streets .
Outcome: The proposed dataset provides textual feedback from machinery visual perception to visually impaired individuals and distracted pedestrians . the results show that the models trained with the dataset can generate meaningful captions focusing on street objects and obstacles in urban scenes .
Multi-Task Semantic Dependency Parsing with Policy Gradient for Learning Easy-First Strategies (P19-1)

Copied to clipboard

Challenge: Existing dependency parsing algorithms do not support directed acyclic graphs . a a systole-based dependency parses sentences using binary semantic relations that are not trees .
Approach: They propose an iterative predicate selection algorithm for semantic dependency parsing . they train the algorithm using multi-task learning and task-specific policy gradient training .
Outcome: The proposed model achieves a new state of the art on the SemEval 2015 task 18 dataset .
Neural Adversarial Training for Semi-supervised Japanese Predicate-argument Structure Analysis (P18-1)

Copied to clipboard

Challenge: Japanese predicate-argument structure analysis involves zero anaphora resolution . state-of-the-art models for PAS analysis achieve an accuracy of around 50% for zero pronouns .
Approach: They propose a Japanese PAS analysis model based on semi-supervised adversarial training with a raw corpus.
Outcome: The proposed model outperforms existing models for Japanese PAS analysis . the model is based on semi-supervised adversarial training with a raw corpus .
ARKitSceneRefer: Text-based Localization of Small Objects in Diverse Real-World 3D Indoor Scenes (2023.findings-emnlp)

Copied to clipboard

Challenge: Existing datasets for 3D referring expression comprehension cover large objects and small objects, such as cooking tools and office supplies.
Approach: They propose a 3D referring expression comprehension dataset that uses 3D scenes to ground text representations onto objects in 3D environments.
Outcome: The proposed dataset covers 15k objects of 1,605 indoor scenes and is significantly larger than existing datasets.
JDocQA: Japanese Document Question Answering Dataset for Generative Language Models (2024.lrec-main)

Copied to clipboard

Challenge: Document question answering is a task of question answering on given documents such as reports, slides, pamphlets, and websites.
Approach: They propose a large-scale document-based QA dataset that requires both visual and textual information to answer questions.
Outcome: The proposed dataset incorporates multiple categories of questions and unanswerable questions from the document for realistic question-answering applications.
Iterative Span Selection: Self-Emergence of Resolving Orders in Semantic Role Labeling (2022.coling-1)

Copied to clipboard

Challenge: Semantic role labeling is the task of labeling semantic arguments for marked semantic predicates.
Approach: They propose a model which combines global decoding and iterative identification for the semantic arguments to consider their roles and relations in the labeling order.
Outcome: The proposed model outperforms existing models in the benchmark datasets of span-based SRL: CoNLL-2005 and CoNll-2012.
Investigating Web Corpus Filtering Methods for Language Model Development in Japanese (2024.naacl-srw)

Copied to clipboard

Challenge: a high quality web corpus is essential for large language models to be developed . strong filtering methods can lead to lesser performance in downstream tasks .
Approach: They build classifiers and language models that can process large amounts of corpora quickly enough for pretraining LLMs.
Outcome: The proposed method is the most accurate and leads to lesser performance in downstream tasks.

What is GenGO?

GenGO is an NLP powered publication search system. It currenctly indexes 30k+ papers from ACL Anthology, and implements multi-aspect summarization, semantic search, and more!

Information

About
Limitations