| 有久 亘 | D, 中間発表 | ソーシャル・コンピューティング | 荒牧 英治, | 安本 慶一, | 若宮 翔子, | PENG SHAOWEN, | 西山 智弘 |
|
title: Comparative Analysis of Japanese Clinical Note Styles Between Physicians and Large Language Models Using Identical Psychiatric Cases: A Quantitative Text Analysis
abstract: Clinical documentation in psychiatry is a language-centered medical practice that captures patients’ narratives, contextual information, and clinicians’ reasoning rather than objective measurements alone. Although large language models (LLMs) have recently been applied to medical text generation, most prior studies focus on accuracy or summarization quality, and it remains unclear whether LLMs can reproduce specialty-specific documentation styles and clinical reasoning. In this study, we constructed standardized psychiatric outpatient scenarios for depression and schizophrenia and collected initial clinical notes written by psychiatrists, internists, and LLMs instructed with specialty roles. Using multidimensional text analysis—including structural similarity, lexical and semantic similarity, topic distribution, lexical diversity, and redundancy metrics—we systematically compared documentation styles across human physicians and LLMs. By focusing on how identical cases are documented by different agents, this work investigates where stylistic and reasoning differences emerge and discusses the implications and limitations of LLM-assisted psychiatric documentation. language of the presentation: Japanese 発表題目: 同一精神科症例における医師と大規模言語モデルのカルテ記載スタイル比較 発表概要: 精神科カルテは、検査値の記録ではなく、患者の語りや文脈、臨床医の解釈を通じた臨床推論を記述する言語的医療文書である。近年、大規模言語モデル（LLM）が医療文書生成に応用され始めているが、多くの研究は正確性や要約品質に焦点を当てており、専門科ごとの記載様式や推論表現の再現性は十分に検証されていない。 本研究では、うつ病および統合失調症の標準化された精神科外来症例を用い、精神科医、内科医、ならびに専門科役割を指定した LLM が作成した初診カルテを収集した。これらのカルテに対して、構造的類似度、語彙・意味的類似度、トピック分布、表現の多様性、冗長性などの多次元テキスト解析を行い、記載スタイルの差異を定量的に比較した。 同一症例に対する記載の違いに着目することで、人間医師と LLM のカルテ記載様式の特徴と限界を明らかにし、精神科領域における LLM 利用の臨床的・教育的含意について検討する。 | |||||||
| XIAO JINGNAN | M, 2回目発表 | ソーシャル・コンピューティング | 荒牧 英治, | 渡辺 太郎, | 若宮 翔子, | PENG SHAOWEN | |
|
title: Reliable LLM-Based Knowledge Graph Enrichment for Recommender Systems via Representation Probing
abstract: Knowledge graphs (KGs) provide structured relational evidence that can benefit explainable and multi-hop recommendation, yet real-world KGs are sparse and incomplete. Large language models (LLMs) can judge or propose candidate KG triples, but their outputs may be unreliable due to hallucination and calibration issues. In this work, I study a controllable screening framework that updates a KG only with reliability-filtered triples. I convert each candidate triple into a fixed-format natural-language statement and obtain an LLM True/False baseline. Using a frozen LLM encoder, I extract statement representations and train a lightweight linear probe (logistic regression) to output a continuous factuality score. A two-threshold rule (raise/veto/keep) then corrects the LLM decisions only when the probe is confident, enabling an explicit precision–recall trade-off via threshold scanning. Experiments on a MovieLens-1M + DBpedia-aligned KG with hard, plausible negatives show that probe-based correction substantially improves test accuracy over the LLM-only baseline (from 0.7427 to 0.8338 with the best setting). As future work, I will evaluate whether the updated KG (KG′) improves downstream KG-aware/GNN recommendation performance using ranking metrics such as HR@K and NDCG@K. language of the presentation: English | |||||||
| RAHMAN MIZANUR | M, 2回目発表 | ディペンダブルシステム学 | 井上 美智子, | 中島 康彦, | 江口 僚太, | 笹田 大翔 |
|
title: Effective Quantized Neural Network on Faulty Memristor Crossbar Array
abstract: Emerging neuromorphic systems based on memristive crossbar arrays offer unprecedented energy efficiency for deep neural networks. However, the practical deployment of quantized neural networks on this hardware is hindered by a synergy of device-level non-idealities: stuck-at faults (SAF), conductance variation, and stochastic noise. While prior research often evaluates these impairments in isolation, this study systematically investigates their combined impact within a unified hardware-aware framework. We assess the efficacy of Fault Injection with Weight Adjustment (FIWA+WA) and Weight Bipartite Matching with Weight Adjustment (WBM+WA). Notably, our framework models stochastic conductance noise as a transient phenomenon applied post-mapping to simulate real-time inference fluctuations. Although the integration of Post-Training Quantization (PTQ) is currently in progress, experimental results using Quantization-Aware Training (QAT) demonstrate that these strategies significantly enhance model resilience. This research confirms that combining weight mapping and adjustment effectively maintains accuracy against hardware faults, proving that both FIWA and WBM+WA are highly effective. language of the presentation: English | ||||||
| ISLAM MD SIHABUL | D, 中間発表 | ディペンダブルシステム学 | 井上 美智子, | 中島 康彦, | 江口 僚太, | 笹田 大翔 |
|
Title: Enhancing the Reliability of Memristor Crossbar-Based Neural Networks Abstract: Memristor crossbar (MC) arrays have emerged as promising hardware accelerators for neural networks (NNs) due to their high energy efficiency and computational performance. However, the immature fabrication process leads to stuck-at faults (SAFs), where memristors become stuck at high or low resistance states, significantly degrading inference accuracy. Our experimental results show that 10% SAFs can cause accuracy drops of up to 60% for multilayer perceptron (MLP) and 50% for AlexNet. To address this challenge, we propose a reliability-aware design framework combining Fault-Injected Weight-Adjusting (FIWA) training and Weighted Bipartite Matching with Weight Adjusting (WBM+WA) mapping. During training, we inject faults and adjust weights to compensate for faulty memristors. During deployment, we use an intelligent mapping algorithm that minimizes sensitivity to faults while allowing weight adjustments. Our experimental results on MLP (MNIST) and AlexNet (CIFAR-10) demonstrate that the proposed method maintains high inference accuracy across fault rates from 5% to 30%, achieving up to 97.61% accuracy for MLP and 85.18% for AlexNet at 30% fault rate. Unlike previous approaches, our method eliminates the need for time-consuming device-specific retraining, making it practical for large-scale deployment of fault-tolerant neuromorphic computing systems. Language of the presentation: English | ||||||