論文

本論文はインターネットを通して提供される様々なサービスに対し，統一的な認証認可の仕組みを実現する汎用認証認可基盤ヤマタノオロチを提案する．汎用の認証認可基盤には“マルチドメイン環境で多様な認証方式をサポートすること”，“サービスによらないアクセス制御を実現すること”，“基盤の負荷を少なくしてスケーラビリティを確保すること”の3つが要求事項としてあげられる．本論文で提案するヤマタノオロチはこれらの要求事項を満たすような認証と認可の機能を提供するシステムである．AAAプロトコルのDiameterと認証フレームワークのEAPを利用することで，ユーザ情報の一元化とマルチドメイン認証を可能とした．またKerberosやXACMLを参考にしたチケットシステムを設計し，サービスによらないアクセス制御とスケーラビリティの確保を実現した．これらの設計に基づいて実装したシステムが複数のインターネットサービスにおいて正しく動作することを検証し，認証および認可の処理時間も実用上問題ない時間で処理が完了することも確認した．

Text-based search engines are widely used to find files in the Internet. However, the users require to understand contexts existing in files and notice appropriate keywords even if the users do not have deep insights of the topic. In addition, text-based search engines require exhaustive search to make keyword indexes. This paper proposes the concept of catalogue that represents relations among files based on grouping and associating by a directed graph. Grouping of files by a catalogue without any text label lets us notice some relations represented by the catalogue. Moreover, a directed graph within a catalogue which associates files can express more detailed relations. We design catalogue system that stores and shares catalogues globally in a distributed and autonomous way. Generalized graph expression of catalogues shared over the Internet bridges any owners, types, and fields of files as long as the files share a same relation. Various services such as digital museums and digital archives are expected to work on top of the shared catalogues.

Text-to-video generation task has witnessed a notable progress, with the generated outcomes reflecting the text prompts with high fidelity and impressive visual qualities. However, current text-to-video generation models are invariably focused on conveying the visual elements of a single scene, and have so far been indifferent to another important potential of the medium, namely a storytelling. In this paper, we examine text-to-video generation from a storytelling perspective, which has been hardly investigated, and make empirical remarks that spotlight the limitations of current text-to-video generation scheme. We also propose an evaluation framework for storytelling aspects of videos, and discuss the potential future directions.

Large language models (LLMs) have demonstrated remarkable performances on a wide range of natural language tasks. Yet, LLMs' successes have been largely restricted to tasks concerning words, sentences, or documents, and it remains questionable how much they understand the minimal units of text, namely characters. In this paper, we examine contemporary LLMs regarding their ability to understand character composition of words, and show that most of them fail to reliably carry out even the simple tasks that can be handled by humans with perfection. We analyze their behaviors with comparison to token level performances, and discuss the potential directions for future research.

Hypergraphs have been attracting more attention because they are suitable for representing real-world networks with complex relations of three or more terms. Clustering hypergraphs is a fundamental task in network analysis. Although several algorithms for finding a community with a theoretical guarantee are already known, almost all require information on the entire hypergraph for calculation. In this study, we propose a personalized PageRank random walk on a hypergraph and an algorithm for hypergraphs to extract a local community, which can be calculated purely locally. We experimentally show that the algorithm outperforms a benchmark method for real-world hypergraphs. We also evaluate the matching rates with the local community and weak ties using F1-scores while varying several parameters for a synthetic hypergraph generated by a stochastic block model.

Community extraction is widely used for various applications concerning network analysis, but its scope has mostly been confined to single-layer network. However, the real-world data are frequently represented as multilayer network, and furthermore, each layer frequently exhibits different properties. It is thus critical to be able to control to what degree each layer should be reflected in the overall network, in order to perform a reliable community extraction. In this paper, we propose a superimposition method called WAPPRS (weighted APPR with restart) that can monotonically increase the degree of reflection and handle a wide range of values for reflection. Our experiments demonstrate that our proposed model outperforms previous models when dealing with the properties desirable for reliable community extraction from multilayer networks.

We demonstrate a field trial of 4K uncompressed streaming over multicast-capable 80 (8λ × 10) Gbps colored optical packet switching network using SOA based broadcast-and-select switch and stacked optical-code label processing. © 2009 OSA.

We are studying the method of developing a platform that delivers large-volume contents efficiently in a many-to-one transmission style, by combining network virtualization technology and the handling of LDGM error correction code. This paper describes its key technologies such as packet routing by using multiple virtual networks, data transmission using redundancy coding, and evaluation of communication qualities, and shows their current outcomes.

Large Language Models (LLMs) have recently been adopted in personalized recommendation systems due to their strong reasoning and language generation capabilities. However, relying on LLMs for core recommendation computation raises concerns regarding latency, cost, transparency, and user controllability. This demo presents a decoupled framework that separates information exploration from information explanation. Recommended items are determined using lightweight graph-based analysis with controllable popularity bias, while a general-purpose LLM is used only for post-hoc interpretation and natural language explanation. The system enables users to iteratively refine recommendation results without retraining models or modifying prompts. A prototype movie recommendation system built on the MovieLens dataset demonstrates that the proposed design achieves low-latency, transparent, and interactive personalization suitable for real-time applications.

While generative AI has accomplished significant advances in visual qualities, generating coherent visual narratives over timeline still remains challenging due to inconsistencies in recurring elements, such as characters and settings, across different frames or scenes. This paper introduces Narrative Graph Prompting, a novel strategy that employs a structured graph representation to enforce visual consistency in AI-generated stories. Our approach defines canonical visual attributes and unique consistency identifiers for key elements, such as characters, settings, and objects, within a central narrative graph. An orchestrating Large Language Model (LLM) then leverages this graph to generate highly detailed, scene-specific prompts for visual generation models, critically injecting these consistent identifiers. This method ensures that visual models render elements with maintained appearance throughout the narrative, reducing the occurrences of visual drift. We demonstrate how this structured approach enhances visual consistency, provides a framework for scalable narrative generation, and offers fine-grained control over story elements, paving the way for more compelling and coherent AI-powered visual storytelling.

Endeavors to calibrate language models to the field of legal text have almost invariably focused on fine-tuning the language models on legal text or scaling up. However, both approaches can be computationally intensive and resource-demanding, limiting their accessibility and scalability. In this paper, we propose a novel method to address these challenges by integrating Graph Attention Networks (GATs) on top of language models. This integration allows for a more effective representation of the dependencies and relationships between different segments of legal documents. Notably, our approach does not require additional fine-tuning on legal-specific datasets or scaling up model sizes. Despite these simplifications, our model achieves improved performance across multiple tasks in the LexGLUE benchmark, demonstrating its effectiveness and efficiency.

Graph summarization is effective when providing a graph to a third party while protecting the detailed topology. Although existing methods emphasize the topological similarity of the summary graph and the original graph, the analysis results of the two graphs are not consistent. In this paper, we propose a method for generating summary graphs to provide consistency in random walk-based analysis. In particular, our method, WalkSum, determines supernodes such that the random walks are unlikely to transition between them and stay within them. It also determines the weight of the superedges, considering the transition probabilities in the original graph. Our evaluations clarified that WalkSum achieves high accuracy over a wide range of summary rates. Compared to one of the state-of-the-art methods, we found that WalkSum is up to 15% more accurate.