| 大槻 優佳 | 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 | |
|
title: Efficient Maintenance of Large-Scale Medical Dictionaries Using Large Language Models
abstract: Dictionaries have been one of the earliest resources utilized in natural language processing, proving useful across a variety of tasks. However, the high costs associated with their construction and maintenance remain a significant challenge. Leveraging manual revision histories of dictionary data to automatically suggest corrections for unedited terms offers a potential solution to ensure quality while reducing maintenance costs. This study proposes a method for automatically correcting metadata in a large-scale medical dictionary containing over 500,000 terms under development. By utilizing large language models, which perform well in zero-shot settings, the system can efficiently estimate various types of information within the dictionary without requiring task-specific configurations. The effectiveness of the proposed approach is demonstrated through experiments that evaluate its performance in resolving variations in gene biomarker expressions, a task requiring specialized medical knowledge. The results suggest that the method can significantly reduce the burden of dictionary maintenance. language of the presentation: Japanese 発表題目: 大規模言語モデルを活用した大規模医療用語辞書メンテナンスの効率化 発表概要: 自然言語処理において，最も早くから利用されてきたリソースが辞書である．辞書は多様なタスクに有用であるが，構築やメンテナンスに要するコストが課題である．我々は，辞書データへの人手の修正履歴を活用し，未修正の用語に対して修正を自動提案するシステムを構築することで，質を担保しつつ低コストでメンテナンスが可能になると考えた．本研究では，構築中の50万語を超える大規模医療辞書において，医療用語のメタデータを自動修正する手法を提案する．実験では，専門的な医学知識を必要とする遺伝子バイオマーカの表記ゆれの修正性能を検証した結果を報告する．本手法は，多くの辞書メンテナンス負担軽減に貢献するものである． | ||||||
| LIU JINGXUAN | M, 2回目発表 | 自然言語処理学 | 渡辺 太郎, | 荒牧 英治, | 上垣外 英剛 | |
|
title: Cross-lingual Scale: An Ignored Role in Multilingual Translation Evaluation
abstract: The rapid progress of Multilingual Neural Machine Translation (MNMT) systems has highlighted the challenge of effectively evaluating their performance. While automatic evaluation metrics are widely used due to the high cost of human evaluation, they often assess MNMT systems' capabilities by averaging scores across different translation directions, raising concerns about fairness. Human evaluation methods like Multidimensional Quality Metrics (MQM) classify and weigh errors, ensuring fairness across languages. However, automatic metrics often assign different scores to translations with identical MQM ratings, indicating language-specific biases. To investigate this issue, we constructed a benchmark by simulating MQM evaluations with GPT-4o, given the limited language coverage of existing MQM datasets. Our analysis shows that most automatic metrics exhibit language biases. To address this, we propose three normalization strategies: (1)language-specific level-wise normalization, (2)language-specific global normalization, and (3)global normalization across languages to mitigate these biases. language of the presentation: English | ||||||
| PUSSEWALA KANKANANGE ASHMARI PRAMODYA | M, 2回目発表 | 自然言語処理学 | 渡辺 太郎, | 荒牧 英治, | 上垣外 英剛, | 坂井 優介 |
|
title: Translating Movie Subtitles by Large Language Models using Movie Meta-information
abstract:Large language models (LLMs) have advanced natural language processing by understanding, generating, and manipulating texts. Although recent studies have shown that prompt engineering can reduce computational effort and potentially improve translation quality, prompt designs specific to different domains remain challenging. Besides, movie subtitle translation is particularly challenging and understudied, as it involves handling colloquial language, preserving cultural nuances, and requires contextual information such as the movie's theme and storyline to ensure accurate meaning. This study aims to fill this gap by focusing on the translation of movie subtitles through the use of prompting strategies that incorporate the movie's meta-information, e.g., movie title, summary, and genre. We build a multilingual dataset which aligns the OpenSubtitles dataset with their corresponding Wikipedia articles and investigate different prompts and their effect on translation performance. Our experiments with GPT-3.5, GPT-4o, and LLaMA-3 models have shown that the presence of meta-information improves translation accuracy. These findings further emphasize the importance of designing appropriate prompts and highlight the potential of LLMs to enhance subtitle translation quality. language of the presentation:English | ||||||
| 坂上 温紀 | M, 2回目発表 | 自然言語処理学 | 渡辺 太郎, | Sakriani Sakti, | 上垣外 英剛, | 坂井 優介 |
|
title: Cross-Lingual Vocabulary Transfer for Low-Resource Languages Using
Bilingual Dictionaries
abstract: Cross-lingual vocabulary transfer plays a promising role in adapting pre-trained language models to new languages, including low-resource languages. Existing approaches that utilize monolingual or parallel corpora face challenges when applied to languages with limited resources. In this work, we propose a simple yet effective vocabulary transfer method that utilizes bilingual dictionaries, which are available for many languages, thanks to descriptive linguists. Our proposed method leverages a property of BPE tokenizers where removing a subword from the vocabulary causes a fallback to shorter subwords. The embeddings of target subwords are estimated iteratively by progressively removing them from the tokenizer. The experimental results show that our approach outperforms existing methods for low-resource languages, demonstrating the effectiveness of a dictionary-based approach for cross-lingual vocabulary transfer. language of the presentation: Japanese 発表題目: 低資源言語のための辞書を用いた言語間語彙転移 | ||||||
| 新宅 慶騎 | M, 2回目発表 | 光メディアインタフェース | 向川 康博, | 清川 清, | 舩冨 卓哉, | 藤村 友貴, | 北野 和哉 |
|
title: Depth Estimation Using Confocal Stereo
abstract: Confocal stereo is a method for depth estimation from images captured using a single camera while sweeping both the lens focus and aperture. This method has the advantage of being able to generate high-resolution absolute depth maps without the need for special hardware. However, it also has some drawbacks, such as the complexity of prior calibration and limitations in applicable scenes due to the assumption of a Lambertian reflection model. Based on the concept of confocal stereo, I aim to develop a method that can be applied to a wider variety of scenes. language of the presentation: Japanese 発表題目: 共焦点ステレオによる深度推定 発表概要: 単一のカメラを用い，レンズのフォーカスと開口をいずれも掃引しながらシーンを撮影して得られた画像から深度推定を行う手法として共焦点ステレオがある．この手法は，特別なハードウェアを用いることなく高解像度な絶対深度マップの生成が可能であるという特長がある一方，事前のキャリブレーションが煩雑であることや，ランバート反射モデルを仮定しているため適用可能なシーンが制限されるという課題がある．共焦点ステレオの考え方を基に，より多くのシーンに適用可能な手法の開発を目指した検討を行う． | |||||||
| AN ZHITING | M, 2回目発表 | サイバネティクス・リアリティ工学 | 清川 清, | 向川 康博, | 内山 英昭, | Perusquia Hernandez Monica, | 平尾 悠太朗 |
| OYEBODE OLUWATOBI OYEWALE | M, 2回目発表 | サイバーレジリエンス構成学 | 門林 雄基, | 林 優一, | 妙中 雄三 | ||
|
title: ADAPTING A CELL PHONE FOR VOICE BIOMETRIC VERIFICATION
abstract: The rise of the mobile era, especially with the availability of the internet, has significantly contributed to the growth of mobile banking, enhancing financial inclusion, particularly among underserved populations. However, reliance on basic mobile devices without advanced biometric security features exposes users to risks of financial fraud, unauthorized account access, and identity theft. This study introduces a lightweight speaker identification system tailored to improve the security of mobile banking in low-resource environments. By utilizing voice biometric authentication, the system confirms the user’s identity through their voice at the end of each transaction, ensuring that only the authorized user can complete the transaction even if their PIN is compromised. The system combines advanced deep learning models, such as Whisper-large-v3 and Emphasized Channel Attention, Propagation, and Aggregation in Time Delay Neural Network (ECAPA-TDNN), to extract detailed voice features, including linguistic patterns and unique speaker traits, ensuring high accuracy and reliability in real-time identification. Additionally, liveness detection is integrated to defend against advanced spoofing, further enhancing system security. Evaluations on real-world data demonstrate the system’s robust performance, achieving 99.59\% accuracy, 99.62\% precision, and an Equal Error Rate (EER) of 0.0017, validating the system’s practicality in constrained environments without requiring extensive computational resources. This solution provides a scalable and secure voice authentication method that is particularly beneficial for underserved communities, advancing both mobile banking security and financial inclusion. language of the presentation: English | |||||||
| ROMERO VICTOR II MILITANTE | D, 中間発表 | ユビキタスコンピューティングシステム | 安本 慶一, | 門林 雄基, | 諏訪 博彦, | 松田 裕貴, | 松井 智一 |
|
title: Sequential Subnetwork Training for Federated Learning in Opportunistic and Infrastructure-Free Settings
abstract: Enabling federated learning in opportunistic settings allows machine learning models to be trained in more challenging environments such as disaster zones, remote regions, and ad hoc deployments. However, conventional federated learning frameworks rely on persistent connections, synchronized training, and full-model exchanges, which are often infeasible in networks where communication is limited to brief, unpredictable device-to-device encounters. To this end, we propose a decentralized framework in which the model is partitioned into non-overlapping components, each assigned to a device and trained locally. Upon encounter, devices exchange components, which are then retrained on new local data. This process proceeds asynchronously and without coordination, avoiding parameter conflicts and reducing communication cost. We evaluate the approach through controlled simulations and real-world deployments on mobile devices, using pre-trained backbones for efficient feature extraction. Results show that the method supports stable training dynamics and maintains competitive performance across a range of data distributions and network topologies. language of the presentation: English | |||||||
| DAMIRAN ZOLBOO | D, 中間発表 | ユビキタスコンピューティングシステム | 安本 慶一, | 加藤 博一, | 諏訪 博彦, | 松井 智一 | |
|
title: A Robust Framework for Yoga Pose Estimation with Contrastive Learning
abstract: In this study, we present a robust and resource-efficient framework for yoga pose estimation that leverages contrastive learning techniques to meet the needs of health and fitness applications. Our approach investigates the potential of SimCLR, MoCo, and BYOL frameworks for improving classification performance, with a particular emphasis on accuracy, computational efficiency, and model resilience. Data augmentation and normalization are incorporated to support generalization and address challenges such as data scarcity and class imbalance. This work demonstrates how contrastive learning can provide practical benefits for yoga pose classification, offering a promising direction for creating scalable and energy-efficient models that adapt to real-world settings. language of the presentation: English | |||||||
| MUHAMMAD FURQAN RASYID | D, 中間発表 | ユビキタスコンピューティングシステム | 安本 慶一, | 清川 清, | 諏訪 博彦, | 松田 裕貴, | 松井 智一 |
|
title: A Real-Time Adaptive System for Sustainable Blinking Behavior
abstract: Prolonged computer use can suppress blinking, causing strain and fatigue. This research proposes a real-time adaptive system that uses adaptive reminders to support healthier blinking behavior, while also exploring the potential of AI-generated suggestions and gamification strategies for driving behavior change. The pilot study focused on the effectiveness of the adaptive reminder component, which significantly increased blink frequency without causing over-blinking (p < 0.0001) and was rated highly for usability (mean = 4.0, SD = 0.71), satisfaction (mean = 4.4, SD = 0.55), and perceived effectiveness (mean = 4.6, SD = 0.55), with participants highlighting the system’s clarity and usefulness. These findings highlight the promise of adaptive reminders and suggest future work is needed to validate the added impact of AI-based suggestions and gamification for promoting long-term behavior change and sustainable blinking habits. language of the presentation: English | |||||||