Safe domain- savvy AI for reliable human–AI synergy

Laboratory on Human-AI Synergy
Senior Lecturer：TRAN Vu Duc

E-mail：
［Research areas］
Artificial Intelligence, Natural Language Processing, Machine Learning
［Keywords］
Deep Learning, Natural Language Understanding, Generative AI, Legal Engineering, Social Media Analysis

Skills and background we are looking for in prospective students

Mathematics (statistics is a bonus), Programing (Python, R, C++), Machine learning (deep learning, optimization). Having experience with Natural Language Processing and Generative AI is a plus point.

What you can expect to learn in this laboratory

Through studying, for instance, generative AI, and idea of blending knowledge in human-understandable form and machine-understandable form to achieve consistent and creative results, students shall obtain knowledge of currently developed methods/technologies and understanding of the methodological evolution of communication with Generative AI systems and be able to utilize available methods/technologies for solving problems and think about new problems and new methods.

【Job category of graduates】AI academia & industry

Research outline

We study topics revolving around generative AI, NLP, and domain-specific applications in, for example, the legal domain, biology, and sociology.

Efficient NLP, Generative AI in specialized domains

Specialized domains require technical understanding often deviated from common understanding, so models trained on common data need adjustment to align with the value required for the given domain, e.g. understanding legal norms in the legal domain, and understanding social norms in social media data. However, for example, re-training is considered costly and environment unfriendly.

For that, it is valuable to make efficient systems from fundamental models for specialized domains. Prompt engineering and retrieval-augmented generation are potential approaches.

Pervasive AI: Understanding complex AI-involved communication networks

Humans and AI are never this close. However, the behaviors of AI agents in communications among themselves and with humans are very complex, and if not well-understood, can cause catastrophic disasters.

Therefore, the ability to analyze the behaviors and predict the behavioral changes can help regulate AI agent network and its operations to ensure its reliability.

Safeguarding AI:

AI is now dominant in daily life though, inability to control its behaviors may lead to severe consequences. Safeguards are being engineered to protect against AI misuses but also are being broken by adversaries. Besides, improving the safety of AI can increase the confidence of utilizing AI in critical domains, e.g. law and healthcare.

It is needed to study more robust methods for safeguarding AI systems from misuses, moving from engineering safeguards with, for instance, dictionary-based methods, to more generalized methods, for example, analyzing semantic space of generative models.

Key publications

1. Tran, V., Le Nguyen, M., Tojo, S., & Satoh, K. (2020). Encoded summarization: summarizing documents into continuous vector space for legal case retrieval. Artificial Intelligence and Law, 28, 441-467.
2. Tran, V., Tran, V. H., Nguyen, P., Nguyen, C., Satoh, K., Matsumoto, Y., & Nguyen, M. (2021, April). CovRelex: A COVID-19 retrieval system with relation extraction. In Proceedings of the 16th conference of the european chapter of the association for computational linguistics: system demonstrations (pp. 24-31).
3. Tran, V., & Matsui, T. (2023). COVID-19 case prediction using emotion trends via Twitter emoji analysis: A case study in Japan. Frontiers in Public Health, 11, 1079315.