Donghee Choi
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
Donghee Choi has a total of 74 co-authors. Here are the top co-authors by number of shared publications:
- Jaewoo Kang
Korea University
7 shared publications - Mogan Gim
Hankuk University of Foreign Studies, Yongin, Republic of Korea
6 shared publications - Jinkyu Kim
Korea University, Seoul, Republic of Korea
3 shared publications - Joram Matthias Posma
Section of Bioinformatics, Division of Systems Medicine, Department of Metabolism, Digestion and Reproduction, Imperial College London, W12 0NN, United Kingdom
3 shared publications - Antoine D. Lain
Section of Bioinformatics, Division of Systems Medicine, Department of Metabolism, Digestion and Reproduction, Imperial College London, W12 0NN, United Kingdom
2 shared publications - Donghyeon Park
Sejong University, Seoul, Republic of Korea
2 shared publications
Latest Publications
- Open Access
- 27 May 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 82812-82822
SciNumQA: A Benchmark for Complex Numerical Reasoning Over Scientific Tables
Numerical reasoning is crucial for extracting meaningful information from scientific tables, enabling deeper analysis beyond simple lookup. As automated analysis of structured data becomes increasingly important not only in scientific research but also in industrial applications such as technical decisionmaking and compliance, the demand for robust numerical reasoning has grown. While recent advances in large language models (LLMs) have improved tabular reasoning, the ability to perform advanced numerical reasoning over scientific tables remains underexplored. Most prior work has focused on simple lookup or basic numerical tasks, often using small or trivial tables. In this work, we introduce SciNumQA, a benchmark dataset comprising 1,607 question–answer pairs based on 572 scientific tables, specifically designed to evaluate complex, multi-step numerical reasoning. Our benchmark systematically filters out trivial cases, ensuring that every question requires compositional reasoning and explicit numerical operations. Experiments show that even recent LLMs (e.g., GPT-4.1 mini) struggle to perform accurate reasoning on these challenging tasks, particularly when table structure becomes more complex. This highlights the need for more robust and generalizable models for scientific table understanding and advanced numerical reasoning. The SciNumQA dataset is available at this link.
- 27 May 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 82812-82822
Open AccessSciNumQA: A Benchmark for Complex Numerical Reasoning Over Scientific Tables
Numerical reasoning is crucial for extracting meaningful information from scientific tables, enabling deeper analysis beyond simple lookup. As automated analysis of structured data becomes increasingly important not only in scientific research but also in industrial applications such as technical decisionmaking and compliance, the demand for robust numerical reasoning has grown. While recent advances in large language models (LLMs) have improved tabular reasoning, the ability to perform advanced numerical reasoning over scientific tables remains underexplored. Most prior work has focused on simple lookup or basic numerical tasks, often using small or trivial tables. In this work, we introduce SciNumQA, a benchmark dataset comprising 1,607 question–answer pairs based on 572 scientific tables, specifically designed to evaluate complex, multi-step numerical reasoning. Our benchmark systematically filters out trivial cases, ensuring that every question requires compositional reasoning and explicit numerical operations. Experiments show that even recent LLMs (e.g., GPT-4.1 mini) struggle to perform accurate reasoning on these challenging tasks, particularly when table structure becomes more complex. This highlights the need for more robust and generalizable models for scientific table understanding and advanced numerical reasoning. The SciNumQA dataset is available at this link.
- Open Access
Relation Extraction for Diet, Non-Communicable Disease and Biomarker Associations (RECoDe): A CoDiet study
Diet plays a critical role in human health, with growing evidence linking dietary habits to disease outcomes. However, extracting structured dietary knowledge from biomedical literature remains challenging due to the lack of dedicated relation extraction datasets. To address this gap, we introduce RECoDe, a novel relation extraction (RE) dataset designed specifically for diet, disease, and related biomedical entities. RECoDe captures a diverse set of relation types, including a broad spectrum of positive association patterns and explicit negative examples, with over 5,000 human-annotated instances validated by up to five independent annotators. Furthermore, we benchmark various natural language processing (NLP) RE models, including BERT-based architectures and enhanced prompting techniques with locally deployed large language models (LLMs) to improve classification performance on underrepresented relation types. The best performing model gpt-oss-20B, a local LLM, achieved an F1-score of 64% for multi-class classification and 92% for binary classification using a hierarchical prompting strategy with a separate reflection step built in. To demonstrate the practical utility of RECoDe, we introduce the Contextual Co-occurrence Summarisation (Co-CoS) framework, which aggregates sentence-level relation extractions into document-level summaries and further integrates evidence across multiple documents. Co-CoS produces effect estimates consistent with established dietary knowledge, demonstrating its validity as a general framework for systematic evidence synthesis. Availability: The code, models, and data will be made freely available upon acceptance.
Open AccessRelation Extraction for Diet, Non-Communicable Disease and Biomarker Associations (RECoDe): A CoDiet study
Diet plays a critical role in human health, with growing evidence linking dietary habits to disease outcomes. However, extracting structured dietary knowledge from biomedical literature remains challenging due to the lack of dedicated relation extraction datasets. To address this gap, we introduce RECoDe, a novel relation extraction (RE) dataset designed specifically for diet, disease, and related biomedical entities. RECoDe captures a diverse set of relation types, including a broad spectrum of positive association patterns and explicit negative examples, with over 5,000 human-annotated instances validated by up to five independent annotators. Furthermore, we benchmark various natural language processing (NLP) RE models, including BERT-based architectures and enhanced prompting techniques with locally deployed large language models (LLMs) to improve classification performance on underrepresented relation types. The best performing model gpt-oss-20B, a local LLM, achieved an F1-score of 64% for multi-class classification and 92% for binary classification using a hierarchical prompting strategy with a separate reflection step built in. To demonstrate the practical utility of RECoDe, we introduce the Contextual Co-occurrence Summarisation (Co-CoS) framework, which aggregates sentence-level relation extractions into document-level summaries and further integrates evidence across multiple documents. Co-CoS produces effect estimates consistent with established dietary knowledge, demonstrating its validity as a general framework for systematic evidence synthesis. Availability: The code, models, and data will be made freely available upon acceptance.
- Open Access
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 140164-140182
ASH: A Systematic Automated Evaluation Framework for Creative Text Generation in Cuisine Transfer Tasks
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 140164-140182
Open AccessASH: A Systematic Automated Evaluation Framework for Creative Text Generation in Cuisine Transfer Tasks
- Open Access
- 10 November 2025
- Published by Association for Computing Machinery (ACM)
- p. 5780-5787
DeepAries: Adaptive Rebalancing Interval Selection for Enhanced Portfolio Selection
- 10 November 2025
- Published by Association for Computing Machinery (ACM)
- p. 5780-5787
Open AccessDeepAries: Adaptive Rebalancing Interval Selection for Enhanced Portfolio Selection
- Open Access
- 16 September 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. PP, 1
RecipeBuild: Learning to Build Recipe Ideas using Language model-based Recipe Contextual Modeling
We introduce RecipeBuild, a fine-tuned Pre-trained Language Model (PLM) specifically trained for enhanced contextual understanding in cooking recipes, including human-written recipe titles and user preferences such as cuisine style and cooking method. Our model excels in two key tasks: the Recipe Completion Task, where it recommends proper ingredients, and the Recipe Ideation Task, where it suggests innovative and diverse ingredient combinations. Experimental results demonstrate that RecipeBuild outperforms the baseline models in the Recipe Completion task, achieving an MRR of 0.0798 and a recall of 0.0794, as well as in the Recipe Ideation task, with a Macro MRR of 0.3606 and a Micro MRR of 0.0641, making it a valuable tool for generating personalized and realistic cooking ideas that cater to various dietary needs and culinary cultures. By addressing the generalized task of ingredient recommendation, RecipeBuild demonstrates the potential for improved recipe representation learning.
- 16 September 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. PP, 1
Open AccessRecipeBuild: Learning to Build Recipe Ideas using Language model-based Recipe Contextual Modeling
We introduce RecipeBuild, a fine-tuned Pre-trained Language Model (PLM) specifically trained for enhanced contextual understanding in cooking recipes, including human-written recipe titles and user preferences such as cuisine style and cooking method. Our model excels in two key tasks: the Recipe Completion Task, where it recommends proper ingredients, and the Recipe Ideation Task, where it suggests innovative and diverse ingredient combinations. Experimental results demonstrate that RecipeBuild outperforms the baseline models in the Recipe Completion task, achieving an MRR of 0.0798 and a recall of 0.0794, as well as in the Recipe Ideation task, with a Macro MRR of 0.3606 and a Micro MRR of 0.0641, making it a valuable tool for generating personalized and realistic cooking ideas that cater to various dietary needs and culinary cultures. By addressing the generalized task of ingredient recommendation, RecipeBuild demonstrates the potential for improved recipe representation learning.
- Open Access
High-precision Biomedical Text Corpora for Multi-Entity Recognition
We present here four biomedical, multi-entity corpora that can be used as benchmarks for named-entity recognition (NER), targeted to literature on metabolic syndrome. The CoDiet-Gold corpus (348,406 annotations) contains 500 re-distributable full-text publications, of which each document was independently annotated by two human experts, with disagreements fully adjudicated by a third expert. The CoDiet-Electrum corpus (3,011,914 annotations) contains 4,445 publications that were annotated using the entities from CoDiet-gold. Finally, for the same 4,445 documents, two fully machine-annotated corpora CoDiet-Bronze (2,952,152 annotations) and CoDiet-Silver (2,309,577 annotations), were created by utilising existing NER algorithms to annotate these. These corpora contain categories (organisms, disease, genes, proteins, metabolites) that add depth to existing corpora, as well as new categories that do not have other corpora (food, dietary methods, sample types, computational methods, study methodology, population characteristics, data types, and microbiome). Availability: These corpora are available at Zenodo ( https://doi.org/10.5281/zenodo.17610205 ).
Open AccessHigh-precision Biomedical Text Corpora for Multi-Entity Recognition
We present here four biomedical, multi-entity corpora that can be used as benchmarks for named-entity recognition (NER), targeted to literature on metabolic syndrome. The CoDiet-Gold corpus (348,406 annotations) contains 500 re-distributable full-text publications, of which each document was independently annotated by two human experts, with disagreements fully adjudicated by a third expert. The CoDiet-Electrum corpus (3,011,914 annotations) contains 4,445 publications that were annotated using the entities from CoDiet-gold. Finally, for the same 4,445 documents, two fully machine-annotated corpora CoDiet-Bronze (2,952,152 annotations) and CoDiet-Silver (2,309,577 annotations), were created by utilising existing NER algorithms to annotate these. These corpora contain categories (organisms, disease, genes, proteins, metabolites) that add depth to existing corpora, as well as new categories that do not have other corpora (food, dietary methods, sample types, computational methods, study methodology, population characteristics, data types, and microbiome). Availability: These corpora are available at Zenodo ( https://doi.org/10.5281/zenodo.17610205 ).
- Open Access
Resource-efficient medical vision language model for dermatology via a synthetic data generation framework
Vision-language models (VLMs), with their ability to integrate visual and textual information, have enabled unified and interpretable multimodal reasoning. However, developing explainable, image-based artificial intelligence (AI) systems for medicine requires locally deployable models designed to ensure privacy-preserving data workflows. Here, we present SCALEMED (Scalable Clinical Assistants and LEarning for MEDicine), a modular framework that enables the development of locally deployable medical VLMs using small models and synthetic data. The SCALEMED framework integrates clinician data annotation, open-source image-text data collection, synthetic data generation through knowledge transfer using larger VLMs, and fine-tuning of small VLMs to develop domain-specific medical AI systems. As a use case in dermatology, we train a resource-efficient VLM, DermatoLlama, which demonstrates higher success rates in report generation compared to state-of-the-art VLMs across text and image-based evaluation datasets. DermatoLlama, based on Llama 3.2, was trained using DermaSynth, a dataset comprising 1.2 million synthetic text samples generated from 367 expert-crafted seed tasks and 82,379 open-source dermatological images. The SCALEMED framework offers a practical solution for developing explainable and accessible medical AI systems, particularly in resource-constrained healthcare environments.
Open AccessResource-efficient medical vision language model for dermatology via a synthetic data generation framework
Vision-language models (VLMs), with their ability to integrate visual and textual information, have enabled unified and interpretable multimodal reasoning. However, developing explainable, image-based artificial intelligence (AI) systems for medicine requires locally deployable models designed to ensure privacy-preserving data workflows. Here, we present SCALEMED (Scalable Clinical Assistants and LEarning for MEDicine), a modular framework that enables the development of locally deployable medical VLMs using small models and synthetic data. The SCALEMED framework integrates clinician data annotation, open-source image-text data collection, synthetic data generation through knowledge transfer using larger VLMs, and fine-tuning of small VLMs to develop domain-specific medical AI systems. As a use case in dermatology, we train a resource-efficient VLM, DermatoLlama, which demonstrates higher success rates in report generation compared to state-of-the-art VLMs across text and image-based evaluation datasets. DermatoLlama, based on Llama 3.2, was trained using DermaSynth, a dataset comprising 1.2 million synthetic text samples generated from 367 expert-crafted seed tasks and 82,379 open-source dermatological images. The SCALEMED framework offers a practical solution for developing explainable and accessible medical AI systems, particularly in resource-constrained healthcare environments.
- Open Access
Culinary Class Wars: Evaluating LLMs using ASH in Cuisine Transfer Task
The advent of Large Language Models (LLMs) have shown promise in various creative domains, including culinary arts. However, many LLMs still struggle to deliver the desired level of culinary creativity, especially when tasked with adapting recipes to meet specific cultural requirements. This study focuses on cuisine transfer-applying elements of one cuisine to another-to assess LLMs' culinary creativity. We employ a diverse set of LLMs to generate and evaluate culturally adapted recipes, comparing their evaluations against LLM and human judgments. We introduce the ASH (authenticity, sensitivity, harmony) benchmark to evaluate LLMs' recipe generation abilities in the cuisine transfer task, assessing their cultural accuracy and creativity in the culinary domain. Our findings reveal crucial insights into both generative and evaluative capabilities of LLMs in the culinary domain, highlighting strengths and limitations in understanding and applying cultural nuances in recipe creation. The code and dataset used in this project will be openly available in \url{http://github.com/dmis-lab/CulinaryASH}.
Open AccessCulinary Class Wars: Evaluating LLMs using ASH in Cuisine Transfer Task
The advent of Large Language Models (LLMs) have shown promise in various creative domains, including culinary arts. However, many LLMs still struggle to deliver the desired level of culinary creativity, especially when tasked with adapting recipes to meet specific cultural requirements. This study focuses on cuisine transfer-applying elements of one cuisine to another-to assess LLMs' culinary creativity. We employ a diverse set of LLMs to generate and evaluate culturally adapted recipes, comparing their evaluations against LLM and human judgments. We introduce the ASH (authenticity, sensitivity, harmony) benchmark to evaluate LLMs' recipe generation abilities in the cuisine transfer task, assessing their cultural accuracy and creativity in the culinary domain. Our findings reveal crucial insights into both generative and evaluative capabilities of LLMs in the culinary domain, highlighting strengths and limitations in understanding and applying cultural nuances in recipe creation. The code and dataset used in this project will be openly available in \url{http://github.com/dmis-lab/CulinaryASH}.
- Open Access
- 21 October 2024
- Published by Association for Computing Machinery (ACM)
- p. 4414-4422
DeepClair: Utilizing Market Forecasts for Effective Portfolio Selection
Utilizing market forecasts is pivotal in optimizing portfolio selection strategies. We introduce DeepClair, a novel framework for portfolio selection. DeepClair leverages a transformer-based time-series forecasting model to predict market trends, facilitating more informed and adaptable portfolio decisions. To integrate the forecasting model into a deep reinforcement learning-driven portfolio selection framework, we introduced a two-step strategy: first, pre-training the time-series model on market data, followed by fine-tuning the portfolio selection architecture using this model. Additionally, we investigated the optimization technique, Low-Rank Adaptation (LoRA), to enhance the pre-trained forecasting model for fine-tuning in investment scenarios. This work bridges market forecasting and portfolio selection, facilitating the advancement of investment strategies.
- 21 October 2024
- Published by Association for Computing Machinery (ACM)
- p. 4414-4422
Open AccessDeepClair: Utilizing Market Forecasts for Effective Portfolio Selection
Utilizing market forecasts is pivotal in optimizing portfolio selection strategies. We introduce DeepClair, a novel framework for portfolio selection. DeepClair leverages a transformer-based time-series forecasting model to predict market trends, facilitating more informed and adaptable portfolio decisions. To integrate the forecasting model into a deep reinforcement learning-driven portfolio selection framework, we introduced a two-step strategy: first, pre-training the time-series model on market data, followed by fine-tuning the portfolio selection architecture using this model. Additionally, we investigated the optimization technique, Low-Rank Adaptation (LoRA), to enhance the pre-trained forecasting model for fine-tuning in investment scenarios. This work bridges market forecasting and portfolio selection, facilitating the advancement of investment strategies.
- Open Access
- 21 October 2024
- Published by Association for Computing Machinery (ACM)
- p. 4637-4644
LAPIS: Language Model-Augmented Police Investigation System
Crime situations are race against time. An AI-assisted criminal investigation system, providing prompt but precise legal counsel is in need for police officers. We introduce LAPIS (Language Model Augmented Police Investigation System), an automated system that assists police officers to perform rational and legal investigative actions. We constructed a finetuning dataset and retrieval knowledgebase specialized in crime investigation legal reasoning task. We extended the dataset's quality by incorporating manual curation efforts done by a group of domain experts. We then finetuned the pretrained weights of a smaller Korean language model to the newly constructed dataset and integrated it with the crime investigation knowledgebase retrieval approach. Experimental results show LAPIS' potential in providing reliable legal guidance for police officers, even better than the proprietary GPT-4 model. Qualitative analysis on the rationales generated by LAPIS demonstrate the model's reasoning ability to leverage the premises and derive legally correct conclusions.
- 21 October 2024
- Published by Association for Computing Machinery (ACM)
- p. 4637-4644
Open AccessLAPIS: Language Model-Augmented Police Investigation System
Crime situations are race against time. An AI-assisted criminal investigation system, providing prompt but precise legal counsel is in need for police officers. We introduce LAPIS (Language Model Augmented Police Investigation System), an automated system that assists police officers to perform rational and legal investigative actions. We constructed a finetuning dataset and retrieval knowledgebase specialized in crime investigation legal reasoning task. We extended the dataset's quality by incorporating manual curation efforts done by a group of domain experts. We then finetuned the pretrained weights of a smaller Korean language model to the newly constructed dataset and integrated it with the crime investigation knowledgebase retrieval approach. Experimental results show LAPIS' potential in providing reliable legal guidance for police officers, even better than the proprietary GPT-4 model. Qualitative analysis on the rationales generated by LAPIS demonstrate the model's reasoning ability to leverage the premises and derive legally correct conclusions.