人工知能学会論文誌
Online ISSN : 1346-8030
Print ISSN : 1346-0714
検索
OR
閲覧
検索
23 巻 , 4 号
選択された号の論文の4件中1~4を表示しています
    • |<
    • <
    • 1
    • >
    • >|
論文
  • 渡邉 陽太郎, 浅原 正幸, 松本 裕治
    23 巻 (2008) 4 号 p. 245-254
    公開日: 2008/04/09
    ジャーナル フリー
    This paper presents a method for categorizing named entities in Wikipedia. In Wikipedia, an anchor text is glossed in a linked HTML text. We formalize named entity categorization as a task of categorizing anchor texts with linked HTML texts which glosses a named entity. Using this representation, we introduce a graph structure in which anchor texts are regarded as nodes. In order to incorporate HTML structure on the graph, three types of cliques are defined based on the HTML tree structure. We propose a method with Conditional Random Fields (CRFs) to categorize the nodes on the graph. Since the defined graph may include cycles, the exact inference of CRFs is computationally expensive. We introduce an approximate inference method using Tree-based Reparameterization (TRP) to reduce computational cost. In experiments, our proposed model obtained significant improvements compare to baseline models that use Support Vector Machines.
    抄録全体を表示
  • 秋山 英久, 野田 五十樹
    23 巻 (2008) 4 号 p. 255-267
    公開日: 2008/04/24
    ジャーナル フリー
    In this paper, we propose a triangulation based function approximation model for agent positioning problem in the dynamic environments. In many problems of the real-world multi-agent/robot domain, a position of each agent is an important factor to affect agents' performance. Because the real-world problem is generally dynamic, a suitable position for each agent should be determined according to the current status of the environment. First, we formalized this issue as a function approximation that maps from state variables to a desirable position of each agent, and proposed a function approximation model using Delaunay triangulation. This method is simple, fast and accurate, so that it can be implemented for real-time and scalable problems. In our previous works, our model showed very high approximation accuracy and good generalization capability for two-dimensional input. However, two-dimensional input is insufficient for more generic problems. Therefore, we extend our previous model so that multi-dimensional input can be taken. The extended model forms tree structure that each node represents a local input space. This structure enables us to maintain the multi-dimensional input space flexibly. The previous model is directly used in each local input space. Therefore, each local input space keeps high accuracy and generalization capability. We implemented the extended model and performed the experiments to evaluate its performance. The result shows our extended model can take the multi-dimensional input adequately.
    抄録全体を表示
  • 岩崎 弘利, 瀬賀 信一郎, 平石 広典, 溝口 文雄
    23 巻 (2008) 4 号 p. 268-280
    公開日: 2008/05/21
    ジャーナル フリー
    In recent years, lots of music content can be stored in mobile computing devices, such as a portable digital music player and a car navigation system. Moreover, various information content like news or traffic information can be acquired always anywhere by a cellular communication and a wireless LAN. However, usability issues arise from the simple interfaces of mobile computing devices. Moreover, retrieving and selecting such content poses safety issues, especially while driving. Thus, it is important for the mobile system to recommend content automatically adapted to user's preference and situation. In this paper, we present the user-adapted program scheduling that generates sequences of content (Program) suiting user's preference and situation based on the Bayesian network and the Constraint Satisfaction Problem (CSP) technique. We also describe the design and evaluation of its realization system, the Personal Program Producer (P3). First, preference such as a genre ratio of content in a program is learned as a Bayesian network model using simple operations such as a skip behavior. A model including each content tends to become large-scale. In order to make it small, we present the model separation method that carries out losslessly compression of the model. Using the model, probabilistic distributions of preference to generate constraints are inferred. Finally satisfying the constraints, a program is produced. This kind of CSP has an issue of which the number of variables is not fixedness. In order to make it variable, we propose a method using metavariables. To evaluate the above methods, we applied them to P3 on a car navigation system. User evaluations helped us clarify that the P3 can produce the program that a user prefers and adapt it to the user.
    抄録全体を表示
訂正記事
    • |<
    • <
    • 1
    • >
    • >|
feedback
Top