(日本語はこちら)
!!News!!
👏 23rd August, 2026: One demo paper accepted to EMNLP2026 System Demonstration. Excited to present in Budapest, Hungary.
👏 16th August, 2026: Three papers accepted to INLG2026. See you in Utrecht.
💪1st August, 2026: I joined Sports AI Tech Lab at CyberAgent as a research scientist. I also continue working at AIST as a visiting researcher.
㊗️ 8th July, 2026: One co-authored paper got SAC Highlight Award at ACL2026
👏 30th April, 2026: One demo paper accepted to IJCAI2026. Congrats Kawamatsu-san!
👏 20th April, 2026: One paper accepted to ACL2026 Industry Track. I will attend the conference in San Diego.
👏 7th April, 2026: One paper accepted to ACL2026 Main Conference. Great work Tomoki Tsujimura!
👏 13rd Feburuary 2026: 4 papers accepted to LREC2026. I will visit Mallorca, Spain.
Affiliations
Research Scientist, Sports AI Tech Lab, CyberAgent, Inc.
Visiting Researcher, Language Information Research Team, Artificial Intelligence Research Center, National Institute of Advanced Industrial Science and Technology (AIST)
Collaboration Opportunities
At CyberAgent’s Sports AI Tech Lab, we are looking for students and research collaborators interested in working with us. Our projects generally involve NLP and/or computer vision applied to sports, but prior expertise in sports AI is not required if you are interested in Sports AI Tech.
For more details, please refer to the below and contact me by email or via X using the details below.
Contact: ishigaki_tatsuya@cyberagent.co.jp, Google Scholar, LinkedIn, GitHub, X
Research
My research focuses on natural language generation (NLG), a subfield of natural language processing. I develop methods for real-world applications, particularly in sports and e-sports, as well as techniques for evaluating these methods rigorously.
I primarily publish at established international conferences in natural language processing, including ACL-affiliated conferences, INLG, SIGDIAL, and LREC. I also publish at international conferences in relevant application domains.
1. Multimodal Language Generation and Real-Time Commentary Generation
This line of research began with generating commentary for racing games from video and structured telemetry data. I subsequently investigated the generation of grounded descriptions from numerical time-series data.
More recently, this work has evolved toward end-to-end, real-time commentary systems for sports and esports. My research now addresses the full commentary-generation pipeline: what to say, when to say it, how to generate it with low latency, and how to transition naturally when an important event occurs during an ongoing utterance.
INLG2021 — pilot work on multimodal racing-game commentary generation
EMNLP2022 — generating grounded descriptions from numerical time-series data
ACL2026 Demo — a live football commentary system
LREC2026 — esports commentary generation
IJCAI2026 — low-latency game commentary through parallel text generation
INLG2026 — mid-utterance transition control for real-time audio game commentary
Together, these studies form the basis of a broader research program on multimodal systems that can understand live events and communicate them to users in a timely, accurate, and engaging manner.
2. Idea Generation with Large Language Models
Another line of my research investigates how large language models can support creative and exploratory thinking.
This work began with the generation of medium- and long-term societal risk scenarios. It then expanded to multi-agent systems for proposing research topics and, more recently, to methods for evaluating business ideas generated or discussed with LLMs.
The long-term objective is to move beyond unconstrained idea generation toward structured, collaborative, and systematically evaluable ideation.
LREC2022 — generating medium- and long-term societal risk scenarios with LLMs
SIGDIAL2026 — research-topic ideation with multiple LLM agents
ACL2026 — evaluating business ideas with LLMs
3. Applications of Large Language Models to Specialized Domains
I also study how LLMs can be applied in domains that require specialized knowledge, constraints, or evaluation criteria.
This research includes evaluating whether LLMs can solve quantum-programming tasks and developing a multimodal system that generates coaching feedback for cyclists from training videos and numerical exercise protocols.
This line is developing from domain-specific capability assessment toward interactive systems that provide useful and grounded support to practitioners and end users.
INLG 2025 — evaluating the quantum-programming capabilities of LLMs
INLG 2026 — generating coaching feedback for cyclists from video and training-protocol data
4. Evaluation and Analysis of Large Language Models
Reliable evaluation is a cross-cutting theme throughout my research.
I investigate not only whether an LLM produces the expected output, but also what capabilities, behavioral patterns, and failure modes underlie its responses. My work in this area has progressed from measuring numerical-sequence understanding to analyzing the consistency and structure of LLMs’ apparent “beliefs.”
INLG 2024 — evaluating LLMs’ ability to understand numerical sequences
ACL 2026 — evaluating and analyzing the apparent “beliefs” of LLMs
Across these four areas, my broader goal is to develop language technologies that are not only technically capable, but also grounded in real-world data, responsive to practical constraints, and evaluated in ways that reflect their actual usefulness.