講演抄録/キーワード |
講演名 |
2012-11-07 15:30
Online Large-margin Weight Learning for First-order Logic-based Abduction ○Naoya Inoue・Kazeto Yamamoto・Yotaro Watanabe・Naoaki Okazaki・Kentaro Inui(Tohoku Univ.) IBISML2012-54 |
抄録 |
(和) |
Abduction is inference to the best explanation. Abduction has long been studied in a wide range of contexts and is widely used for modeling artificial intelligence systems, such as diagnostic systems and plan recognition systems. However, less attention has been paid to how to automatically learn score functions, which rank explanations in order of their plausibility. In this paper, we propose a supervised learning approach for first-order logic-based abduction. The contribution of this paper is the following: (i) we show how to formulate the machine learning problem of abduction with the framework of online large-margin training, which has been shown to have both predictive performance and scalability to larger problems; (ii) we extend the state-of-the-art abductive reasoning system to model the score function with a weighted linear model, which is the groundwork for the online large-margin training; (iii) we support partially-specified gold-standard explanations as training examples, where the weights are learned to rank any explanation that includes the gold-standard explanation as the best explanation; (iv) the all-in-one software package for inference and learning is made publicly available. |
(英) |
Abduction is inference to the best explanation. Abduction has long been studied in a wide range of contexts and is widely used for modeling artificial intelligence systems, such as diagnostic systems and plan recognition systems. However, less attention has been paid to how to automatically learn score functions, which rank explanations in order of their plausibility. In this paper, we propose a supervised learning approach for first-order logic-based abduction. The contribution of this paper is the following: (i) we show how to formulate the machine learning problem of abduction with the framework of online large-margin training, which has been shown to have both predictive performance and scalability to larger problems; (ii) we extend the state-of-the-art abductive reasoning system to model the score function with a weighted linear model, which is the groundwork for the online large-margin training; (iii) we support partially-specified gold-standard explanations as training examples, where the weights are learned to rank any explanation that includes the gold-standard explanation as the best explanation; (iv) the all-in-one software package for inference and learning is made publicly available. |
キーワード |
(和) |
アブダクション / 論理推論 / オンライン学習 / ラージマージン学習 / 構造学習 / 隠れ変数 / Passive Aggressive アルゴリズム / |
(英) |
Abduction / Logic-based Reasoning / Online learning / Large-margin training / Structured learning / Latent variables / Passive Aggressive algorithm / |
文献情報 |
信学技報, vol. 112, no. 279, IBISML2012-54, pp. 143-150, 2012年11月. |
資料番号 |
IBISML2012-54 |
発行日 |
2012-10-31 (IBISML) |
ISSN |
Print edition: ISSN 0913-5685 Online edition: ISSN 2432-6380 |
著作権に ついて |
技術研究報告に掲載された論文の著作権は電子情報通信学会に帰属します.(許諾番号:10GA0019/12GB0052/13GB0056/17GB0034/18GB0034) |
PDFダウンロード |
IBISML2012-54 |
|