人工知能学会論文誌
Online ISSN : 1346-8030
Print ISSN : 1346-0714
ISSN-L : 1346-0714
最新号
選択された号の論文の10件中1~10を表示しています
原著論文
  • 大塚 淳史, 西田 京介, 斉藤 いつみ, 浅野 久子, 富田 準二, 佐藤 哲司
    原稿種別: 原著論文
    2019 年 34 巻 5 号 p. A-J14_1-12
    発行日: 2019/09/01
    公開日: 2019/09/01
    ジャーナル フリー

    The performance of reading comprehension, which is a question answering technique, by deep neural networks is now comparable to that of humans. However, there are still problems with the reading comprehension when given ambiguous questions. In this work, we propose a novel task called Specific Question Generation (SQG). SQG specifically revises the input question and suggests several specific question (SQ) candidates so that users can choose the SQ that is closest to their intent and obtain a highly accurate answer from the reading comprehension. We also propose a Specific Question Generation Model (SQGM) for facilitating the SQG. This model is based on an encoder-decoder model and uses two copy mechanisms (question copy and passage copy). The key idea here is that the missing information in the user-input question is described in the passage. Experimental results with public reading comprehension datasets demonstrated that our model generated specific questions that can improve reading comprehension accuracy.

  • 立花 誠人, 村田 剛志
    原稿種別: 原著論文
    2019 年 34 巻 5 号 p. B-IC2_1-8
    発行日: 2019/09/01
    公開日: 2019/09/01
    ジャーナル フリー

    Since several types of data can be represented as graphs, there has been a demand for generalizing neural network models for graph data. Graph convolution is a recent scalable method for performing deep feature learning on attributed graphs by aggregating local node information over multiple layers. Such layers only consider attribute information of node neighbors in the forward model and do not incorporate knowledge of global network structure in the learning task. In this paper, we present a scalable semi-supervised learning method for graph-structured data which considers not only neighbors information, but also the global network structure. In our method, we add a term preserving the network structural features such as centrality to the objective function of Graph Convolutional Network and train for both node classification and network structure preservation simultaneously. Experimental results showed that our method outperforms state-of-the-art baselines for the node classification tasks in the sparse label regime.

  • 北村 拓也, 平嶋 宗
    原稿種別: 速報論文
    2019 年 34 巻 5 号 p. C-I73_1-5
    発行日: 2019/09/01
    公開日: 2019/09/01
    ジャーナル フリー

    Regarding cultivation of the capacity to think logically, we designed/developed system that enables segmentation task and measured time change and learning effect by application order using structuring task system developed in the existing study. Subjects were 14 university students/graduate students. In order to verify the difference of task difficulty by exercise working order, we analysed exercise time. As a result, the structuring task first group showed significant decrease of segmentation task exercise time compared to that of the segmentation first group.

  • Daigo Okada, Naotoshi Nakamura, Takuya Wada, Ayako Iwasaki, Ryo Yamada
    原稿種別: Original Paper
    2019 年 34 巻 5 号 p. D-J13_1-7
    発行日: 2019/09/01
    公開日: 2019/09/01
    ジャーナル フリー

    In the field of bioimaging, an important part of analyzing the motion of objects is tracking. We propose a method that applies the Sinkhorn distance for solving the optimal transport problem to track objects. The advantage of this method is that it can flexibly incorporate various assumptions in tracking as a cost matrix. First, we extend the Sinkhorn distance from two dimensions to three dimensions. Using this three-dimensional distance, we compare the performance of two types of tracking technique, namely tracking that associates objects that are close to each other, which conventionally uses the nearest-neighbor method, and tracking that assumes that the object is moving at constant velocity, using three types of simulation data. The results suggest that when tracking objects moving at constant velocity, our method is superior to conventional nearest-neighbor tracking as long as the added noise is not excessively large. We show that the Sinkhorn method can be applied effectively to object tracking. Our simulation data analysis suggests that when objects are moving at constant velocity, our method, which sets acceleration as a cost, outperforms the traditional nearest-neighbor method in terms of tracking objects. To apply the proposed method to real bioimaging data, it is necessary to set an appropriate cost indicator based on the movement features.

  • 小柴 等, 森川 想
    原稿種別: 原著論文
    2019 年 34 巻 5 号 p. E-J47_1-10
    発行日: 2019/09/01
    公開日: 2019/09/01
    ジャーナル フリー

    In this study, we attempt to reveal the relationship between the National Diet of Japan and Japanese ministries by text analysis of minutes data. The policy making process mainly consists of two routes: One is the parliamentary initiative route, and the other is the ministries initiative route, which is often consulted by advisory committees. These policy making process routes are not independent, but affect one another. While there are many studies and reports that have explored these relationships, most of them are qualitative case studies, which have some methodological limitations such as little comparability among cases.

    We propose a method of measuring the relationship between the National Diet of Japan and Japanese ministries through text similarity and time stamps contained within minutes of public organizations, which have been published online, providing machine-readable open data. Our analysis suggests that the method draws consistent results with existing qualitative analyses and can effectively support and improve understanding of the relationship between the Diet and ministries. In addition, this method has an advantage of analyzing a wide variety of topics using the same method, ensuring comparability for researchers.

  • 是枝 祐太, 間瀬 久雄, 柳井 孝介
    原稿種別: 原著論文
    2019 年 34 巻 5 号 p. F-IC1_1-11
    発行日: 2019/09/01
    公開日: 2019/09/01
    ジャーナル フリー

    Japan Patent Office manually annotates submitted patents with F-terms (a patent classification scheme consisting of more than 300,000 labels) to aid search for prior patent applications. Keeping up the quality of F-term annotation is critical to patentability assessments, thus there is a demand for an automatic way to assist F-term annotation. One potential solution is to point out annotation mistakes by utilizing machine learning-based classification. However, the annotators cannot validate the predicted corrections because conventional classification methods do not give the rationales behind the corrections. Thus, the annotators may only adopt all or no corrections. The goal of this study was to assist F-term annotation by presenting annotators with corrections on the F-term annotation and the rationales behind the corrections.

    We proposed a joint neural model for F-term annotation and rationale identification. The proposed method incorporates a large portion of data annotated only with F-terms and a small portion of data annotated with rationales. It was first trained for F-term annotation, and then fine-tuned using the ground-truth rationales to discriminate rationales from non-rationales.

    We evaluated the proposed method on multiple F-terms from different technical domains. The proposed method outperformed baseline methods in terms of the rationale identification, implying that incorporating rationales in training is particularly useful in identifying rationales.

特集論文「知能創発とネットワーク」
  • 高野 海斗, 酒井 浩之, 北島 良三
    原稿種別: 特集論文
    2019 年 34 巻 5 号 p. wd-A_1-22
    発行日: 2019/09/01
    公開日: 2019/09/01
    ジャーナル フリー

    In this research, we propose a method of extracting business segments from securities reports and extracting sentences containing causal and result information concerning business performance for each extracted business segments. For example, our method extracts“ In the aluminum rolled products business, shipments of high purity foils for aluminum electrolytic capacitors for industrial equipment and automotive use increased and sales increased. ”as a sentence containing causal information concerning business performance. Moreover, our method estimates that the sentence belongs to business segment“ aluminum ”. Our method extracts“ As a result of this segment sales in this segment were 105,439 million yen, operating profit was 6,697 million yen. ”as a sentences containing performance result information belong to“aluminum”segment. We evaluated our method and the method of extracting sentences containing causal information attained 69.3% precision and 72.5% recall, and the method of extracting sentences containing performance result information attained 78.8% precision and 91.1% recall.

  • 清水 仁, 松林 達史, 納谷 太, 澤田 宏
    原稿種別: 特集論文
    2019 年 34 巻 5 号 p. wd-B_1-8
    発行日: 2019/09/01
    公開日: 2019/09/01
    ジャーナル フリー

    The theme park problem is a platform where methods of guiding visitors to relieve congestion are developed and evaluated by reproducing a crowded theme park on a computer. In the theme park problem, the attraction selection model is an important element in the simulator. In previous studies, multinomial logit model was mainly used for attraction selection. However, when we observed real amusement parks, we found that we can not reproduce the characteristics of waiting time of real attraction by this model. In this research, we propose a multinomial linear model as a model of attraction selection. This model can express the rational behavior of visitors that waits for a while when waiting times of all attractions are too long for them. We showed that this model can reproduce characteristics of waiting time using multiagent simulator (MAS). We also developed a method to estimate the parameters of the proposed model from the aggregated data of the output of MAS. As a result of numerical experiments, it was confirmed that the performance of the parameter estimation was good. The proposed model and method for parameter estimation can be applied not only to the theme park problem but also to various problems related to human behavior of selection.

  • 工藤 航, 鳥海 不二夫
    原稿種別: 特集論文
    2019 年 34 巻 5 号 p. wd-E_1-9
    発行日: 2019/09/01
    公開日: 2019/09/01
    ジャーナル フリー

    With the spread of social media and e-commerce websites, the technology of user profiling from users’ action history has attracted a lot of interest. Users’ action history can be acquired both in the passive way and in the active (interactive) way, and previous studies have found out how to presume the users’ profile from their action history which is acquired passively. The purpose of this study was to find out the best way to interact with users for efficient user profiling. First, we constructed a CNN (Convolutional Neural Network) model which presumes user’s profile. Next, we proposed three ways to interact with users for efficient user profiling and compared them. Consequently, two ways of the three were discovered to be effective to streamline the user profiling. And we provided some analyses of the ways, which revealed that the CNN model’s output for an item could be utilized for efficient user profiling , and considering the probability that each item is consumed by the target user could provide more efficient ways.

  • 松林 達史, 清武 寛, 幸島 匡宏, 戸田 浩之, 田中 悠介, 六藤 雄一, 塩原 寿子, 宮本 勝, 清水 仁, 大塚 琢馬, 岩田 ...
    原稿種別: 特集論文
    2019 年 34 巻 5 号 p. wd-F_1-11
    発行日: 2019/09/01
    公開日: 2019/09/01
    ジャーナル フリー

    Forming security plans for crowd navigation is essential to ensure safety management at large-scale events. The Multi Agent Simulator (MAS) is widely used for preparing security plans that will guide responses to sudden and unexpected accidents at large events. For forming security plans, it is necessary that we simulate crowd behaviors which reflects the real world situations. However, the crowd behavior situations require the OD information (departure time, place of Origin, and Destination) of each agent. Moreover, from the viewpoint of protection of personal information, it is difficult to observe the whole trajectories of all pedestrians around the event area. Therefore, the OD information should be estimated from the several observed data which is counted the number of passed people at the fixed points.

    In this paper, we propose a new method for estimating the OD information which has following two features. Firstly, by using Bayesian optimization (BO) which is widely used to find optimal hyper parameters in the machine learning fields, the OD information are estimated efficiently. Secondly, by dividing the time window and considering the time delay due to observation points that are separated, we propose a more accurate objective function.

    We experiment the proposed method to the projection-mapping event (YOYOGI CANDLE 2020), and evaluate the reproduction of the people flow on MAS. We also show an example of the processing for making a guidance plan to reduce crowd congestion by using MAS.

feedback
Top