NTCIR-18

Introduction

NTCIR-18 (The 18th NII Testbeds and Community for Information access Research) is a biennial evaluation workshop series organized by the National Institute of Informatics (NII) to advance research in information access technologies such as information retrieval, question answering, and text summarization. Held from June 10–13, 2025 at NII in Tokyo, Japan, the event gathered international researchers and featured tasks including MedNLP-CHAT and FinArg-2. The IMNTPU team from National Taipei University’s Department of Information Management, led by Professor Min-Yuh Day, participated in this prestigious conference and contributed to discussions on multilingual medical AI and financial argument inference.

Tasks

Our team participated in two main tasks during NTCIR-18:

MedNLP-CHAT

A medical natural language processing task focused on developing conversational AI systems for healthcare applications. This task involved understanding medical terminology, patient queries, and providing appropriate responses in clinical contexts.

FinArg-2

A financial argument mining task that challenges participants to identify, extract, and analyze argumentative structures in financial documents. This task focuses on understanding financial reasoning and decision-making processes.

Methodology

Our research employed innovative methodologies for both tasks, combining state-of-the-art approaches in natural language processing and machine learning.

🏥

MedNLP-CHAT Methodology

🤖

Agentic AI Framework

Multi-model collaborative approach with majority voting and weighted scoring systems

💬

Prompt Engineering

Three-shot prompting with GPT-4o, Claude 3.5 Sonnet, Mistral small latest

⚖️

Risk Assessment

Medical, legal, and ethical risk evaluation across multiple languages

💰

FinArg-2 Methodology

🔧

Model Fine-tuning

BERT, RoBERTa, DistilBERT for encoder-based tasks; GPT-4o Mini for decoder-based

📊

Data Augmentation

Semantic variation generation to address class imbalance in financial datasets

⏰

Temporal Reasoning

Classification of argument claims based on temporal references in financial contexts

Results

Our team achieved significant results in both tasks, demonstrating innovative approaches to natural language processing in specialized domains. The detailed results and methodologies will be published in the conference proceedings.