GAO ZHIWEI | D, 中間発表 | ソーシャル・コンピューティング | 荒牧 英治, | Sakriani Sakti, | 若宮 翔子, | PENG SHAOWEN | |
title: From Cultural Adaptation to Norm Alignment: Probing and Improving LLMs’ Reasoning over Social Norms
abstract: Cultural alignment is increasingly recognized as a critical aspect of deploying large language models (LLMs) in diverse sociocultural settings. In our previous work, we proposed a dual-method framework combining emotion analysis via valence–arousal–dominance (VAD) modeling and a Hofstede-based questionnaire to evaluate the cultural adaptability of multilingual LLMs in Japanese workplace contexts. Our findings revealed substantial differences in how models reflected culturally expected emotional tones, with LLM-jp and Qwen exhibiting stronger alignment with hierarchical and collectivist norms. However, we also observed that affective conformity does not necessarily imply deeper normative understanding. Extending this line of inquiry, our current work investigates whether norm alignment—the ability of LLMs to evaluate the acceptability of social and moral justifications—can be improved through inference-time intervention (ITI). Motivated by the high cost of reinforcement learning with human feedback (RLHF), we explore ITI as a lightweight, controllable alternative. While this study is ongoing, we have demonstrated the feasibility of ITI for steering model behavior on norm-sensitive tasks, suggesting a promising direction for scalable alignment without retraining. language of the presentation: English | |||||||
PIERRE JUDE CRENER JUNIOR | D, 中間発表 | ソーシャル・コンピューティング | 荒牧 英治, | Sakriani Sakti, | 若宮 翔子, | PENG SHAOWEN | |
title: Persona-Based Data Augmentation for Rare Disease Named Entity Recognition abstract: Rare disease Named Entity Recognition (NER) suffers from severe data scarcity due to costly expert annotations and highly specialized terminology. While Large Language Models (LLMs) offer promising advantages for synthetic data generation, augmentation methods using LLMs often struggle with factual inaccuracies and limited lexical diversity. This research explores a novel persona-based data augmentation strategy for rare disease NER, by using controlled prompt engineering to balance linguistic variation and factual accuracy. Through systematic evaluation, we assessed the impact of each persona on both data diversity, fidelity, and model performance. Our findings show that persona-based augmentation improves NER performance over human-annotated baselines.This work demonstrates the potential of persona-driven generative AI for addressing data scarcity in specialized biomedical NLP tasks. language of the presentation: English | |||||||
BENITA ANGELA TITALIM | D, 中間発表 | ヒューマンAIインタラクション | Sakriani Sakti, | 荒牧 英治, | 大内 啓樹, | Faisal Mehmood, | Bagus Tris Atmaja |
title: Adaptive Generative AI for Personalized Hearing Rehabilitation
abstract: Millions of people with hearing loss struggle to understand speech in noisy settings, where traditional hearing aids and enhancement methods often fail. This research addresses this by developing a real-time adaptive system powered by generative AI, combining personalized acoustic adjustments with intelligent language modeling. Focusing on automatic speech recognition (ASR), currently this work integrate large language models (LLMs) through in-context learning to improve recognition in challenging conditions—such as background music with vocals or irrelevant speech—without retraining or altering the ASR architecture. This approach enhances speech intelligibility and paves the way for more effective, user-aware hearing assistive technologies. In the future, this work envisions incorporating personalized text-to-speech (TTS) systems that re-synthesize ASR outputs in a listener-aware manner—adjusting speaking rate, volume, or linguistic complexity—to further enhance comprehension for individuals with hearing impairments. This approach aims to create a closed-loop, user-centric speech interface that adapts both recognition and reproduction for optimal intelligibility and accessibility. language of the presentation: English | |||||||