人工知能学会論文誌
Online ISSN : 1346-8030
Print ISSN : 1346-0714
ISSN-L : 1346-0714
36 巻 , 1 号
選択された号の論文の11件中1~11を表示しています
一般論文
原著論文
  • 井田 安俊, 藤原 靖宏, 鹿島 久嗣
    原稿種別: 原著論文
    2021 年 36 巻 1 号 p. A-JB1_1-11
    発行日: 2021/01/01
    公開日: 2021/01/01
    ジャーナル フリー

    Sparse Group Lasso is a method of linear regression analysis that finds sparse parameters in terms of both feature groups and individual features. Block Coordinate Descent is a standard approach to obtain the parameters of Sparse Group Lasso, and iteratively updates the parameters for each parameter group. However, as an update of only one parameter group depends on all the parameter groups or data points, the computation cost is high when the number of the parameters or data points is large. This paper proposes a fast Block Coordinate Descent for Sparse Group Lasso. It efficiently skips the updates of the groups whose parameters must be zeros by using the parameters in one group. In addition, it preferentially updates parameters in a candidate group set, which contains groups whose parameters must not be zeros. Theoretically, our approach guarantees the same results as the original Block Coordinate Descent. Experiments show that our method enhances the efficiency of original method and achieves the same prediction error.

  • 微分可能メモリを用いた方策オフ型方策勾配法の安定化手法
    妹尾 卓磨, 今井 倫太
    原稿種別: 原著論文
    2021 年 36 巻 1 号 p. B-K71_1-8
    発行日: 2021/01/01
    公開日: 2021/01/01
    ジャーナル フリー

    Deep reinforcement learning has been investigated in high-dimensional continuous control tasks. Deep Deterministic Policy Gradients (DDPG) is known as a highly sample-efficient policy gradients algorithm. However, it is reported that DDPG is unstable during training due to bias and variance problems of learning its action-value function. In this paper, we propose Policy Gradients with Memory Augmented Critic (PGMAC) that builds action-value function with the memory module previously proposed as Differentiable Neural Dictionary (DND). Although the DND is only studied in discrete action-space problems, we propose Action-Concatenated Key, which is a technique to combine DDPG-based policy gradient methods and DND. Furthermore, the remarkable advantage of PGMAC is shown that long-term reward calculation and weighted summation of value estimation at DND has an essential mechanism to solve the bias and variance problem. In experiment, PGMAC significantly outperformed baselines in continuous control tasks. The effects of hyperparameters were also investigated to show that the memory-augmented action-value function reduces the bias and variance in policy optimization.

特集論文「Webインテリジェンスとインタラクション」
原著論文
  • 山西 良典, 藤岡 寛子, 西原 陽子
    原稿種別: 特集論文
    2021 年 36 巻 1 号 p. WI2-A_1-8
    発行日: 2021/01/01
    公開日: 2021/01/01
    ジャーナル フリー

    This paper proposes a method to classify the aspects of restaurant reviews with the pseudo corpus. Restaurant reviews include some types of aspects such as meals and hospitality, which are the target in this paper. Those two types of aspects are important to evaluate the restaurant from review reading. The opinions for each aspect are so useful for the people to know how the restaurant is. However, freely written reviews often do not organize opinions without the classification of the aspects. In the proposed method, the reviews are automatically classified into two types without any annotated review data. The proposed method applies the existing corpora as the pseudo corpus for the target aspects of restaurant reviews. Through the experiment, it was confirmed that the proposed method classified the aspects of restaurant reviews as 71% accuracy and 70% AUC for sentences and 86% accuracy and 93% AUC for reviews. Even though the proposed method did not learn the annotated restaurant reviews, it showed a certain level of effectiveness for the classification better than the chance level.

  • 橋口 友哉, 山本 岳洋, 藤田 澄男, 大島 裕明
    原稿種別: 特集論文
    2021 年 36 巻 1 号 p. WI2-B_1-13
    発行日: 2021/01/01
    公開日: 2021/01/01
    ジャーナル フリー

    In this study, we tackle the problem of retrieving questions from a corpus archived in a Community Question Answering service that a consultant having distress can feel empathy with them. We hypothesize that the consultant feels empathy with the questions having a similar situation with that of the consultant’s distress, and propose a method of retrieving similar sentences focusing on the situation of the distress. Specifically, we propose two approaches to fine-tuning the pre-trained BERT model so that the learned model better captures the similarity of the situation between distress. One tries to extract only the words representing the situation of the distress, the other tries to predict whether the two sentences show the same situation. The data for training the models are gathered by the crowdsourcing task where the workers are asked to gather the sentences whose situation is similar to the given sentence and to annotate the words in the sentences that represent the situation. The data is then used to fine-tune the BERT model. The effectiveness of the proposed methods is evaluated with the baselines such as TF-IDF, Okapi BM25, and the pre-trained BERT. The results of the experiment with 20 queries showed that one of our methods achieved the highest nDCG@5 while we could not observe any significant differences among the methods.

  • 中田 祐誠, 村本 直樹, 山本 岳洋, 藤田 澄男, 大島 裕明
    原稿種別: 特集論文
    2021 年 36 巻 1 号 p. WI2-C_1-10
    発行日: 2021/01/01
    公開日: 2021/01/01
    ジャーナル フリー

    In this study, we propose a method to predict whether a web searcher will purchase a camera in a near future based on his/her web search log. With the increasing popularity of online shopping at EC sites, more and more users are searching for products through web searches and actually purchasing them at EC sites. This indicates that, by analyzing the query log of a searcher, it is possible to predict whether the searcher will purchase the product in the near future. Therefore, we construct a classifier by collecting past web search query logs of searchers who have purchased cameras and those who have not purchased them. In the experiment, we used a web search query log of Yahoo! JAPAN and the product purchase histories of Yahoo! JAPAN Shopping to verify the results. We collected thousands of users who purchased cameras in a certain period and other users in the same number who didn’t purchase but issued queries related to cameras. By analyzing the classifier trained with the prepared dataset, we verify the accuracy of the prediction, the period of time required for the prediction, and whether there are any characteristic words that suggest the purchase.

  • Yasufumi Takama, Jing-cheng Zhang, Hiroki Shibata
    原稿種別: Special Paper
    2021 年 36 巻 1 号 p. WI2-D_1-10
    発行日: 2021/01/01
    公開日: 2021/01/01
    ジャーナル フリー

    This paper proposes a method for recommending music items without explicit feedback. Context and content features are used as auxiliary information to compensate for implicit feedback. The recent development of communication technology and portable electronic devices has changed the way of consuming music. We can access a vast amount of music items via online music streaming services. As a result, finding appropriate music items from enormous resources gets to be difficult for users. To help users discover their favorite music items, recommender systems in the music domain have been studied. This paper focuses on two challenges specific to music recommender systems: the difficulty of obtaining explicit feedback such as ratings, and the importance of making use of context information. To handle the context information as auxiliary information to compensate for implicit feedback, this paper employs FMs (Factorization Machines), in which the context information is treated as features. Utilizing the merit of FMs that can easily introduce features, this paper also introduces content features in addition to context features. As it is known that using low-level content features directly is not effective because of the semantic gap, this paper proposes two types of abstract content features: UGP (user genre profile) and UCP (user content profile). The effectiveness of the proposed method and the effect of negative sampling methods are evaluated in terms of MPR with #nowplaying-rs and LFM-1b dataset. The result of the experiment shows that the proposed method outperforms wALS (weighted Alternating Least Squares), which is one of the popular music recommendation algorithms based on matrix factorization. The characteristics of the proposed sampling methods are investigated with different settings of the parameters and the ratio of negative samples. As for the effectiveness of each feature, it is found that a feature is effective when JS (Jensen-Shannon) divergence of popularity distribution among different feature values is large. It is also shown that the UCP and UGP cluster labels are more effective than using content features directly.

  • 西 朋里, 小川 祐樹, 高 史明, 高野 雅典, 森下 壮一郎, 服部 宏充
    原稿種別: 特集論文
    2021 年 36 巻 1 号 p. WI2-E_1-9
    発行日: 2021/01/01
    公開日: 2021/01/01
    ジャーナル フリー

    With the rise of Internet TV and other new media, people are now viewing the news through a variety of conduits. In addition, the influence of news media on people is changing. Viewers can post comments in Internet TV, and these comments has the viewers’ opinions of the news contents. Therefore, analysis of viewers comments is important in revealing the effect of the news. In addition, these comments are posted based on the morality of the viewers, and point of view of morality is considered important in the analysis of viewer comments in news. Therefore, this study purpose is to clarify the opinion on Internet TV news programs from a moral-based analysis viewers’ comments. This study analyzed the trend of viewer comments on ABEMA news programs using comment length and the application of two methods. First, the morality of viewer comments was analyzed by calculating the moral/immoral expression rate for each program using the moral foundation dictionary. Second, the distributed expressions of viewer comments (calculated by Doc2Vec) were clustered by k-means++, and program trends were analyzed using the cluster characteristics. The results indicated that there was no difference in comment length between the two program types. Comments on soft-news programs had a high moral/immoral expression rate for politics or current events. In contrast, comments of hard-news programs did not show a characteristic trend. A viewer can easily participate in the discussion, because the soft-news program deals with the same news for a long time as the news content is limited compared to the non-discussion program.

  • 江上 周作, 呂 暁東, 古賀 禎, 住谷 泰人
    原稿種別: 特集論文
    2021 年 36 巻 1 号 p. WI2-F_1-12
    発行日: 2021/01/01
    公開日: 2021/01/01
    ジャーナル フリー

    With the advancement of information and communication technologies, to improve the interoperability between heterogeneous information systems by regularizing the syntax for information exchange is essential to achieve global seamless air traffic management operation. However, the current point-to-point aviation related information exchange among different systems and operators cannot satisfy the requirement for interoperability. The concept of System Wide Information Management (SWIM) has been promoted by the International Civil Aviation Organization (ICAO) to implement interoperability and harmonization in a global operation. In the SWIM environment, all the related stakeholders need to efficiently obtain the necessary ATM data with situational awareness from various information domains. However, this is difficult to realize in the current system, as different data are structured based on different data models. In this study, we construct domain ontologies based on flight, aeronautical, and weather information exchange models. Moreover, for semantic interoperability in the SWIM environment, we develop an upper ontology-based reference ontology that enables common situational awareness of spatiotemporal concepts. Furthermore, we propose a methodology for mapping heterogeneous domain ontologies to the reference ontology with the manual refinement. Finally, we apply the proposed ontologies to a SWIM test system. The applicability and scalability of the proposed ontologies are demonstrated through a case study in the SWIM environment.

  • 高野 海斗, 酒井 浩之, 中川 慧
    原稿種別: 特集論文
    2021 年 36 巻 1 号 p. WI2-G_1-19
    発行日: 2021/01/01
    公開日: 2021/01/01
    ジャーナル フリー

    A shareholder convocation notice is a letter that the company is obliged to send to shareholders when holding a shareholder’s meeting. We can access it on corporate websites and acquire as PDF file. It contains a lot of useful information, such as company profile, major shareholders, and bills to be discussed. Therefore, institutional investors often use that information in their investment decisions. However, the following challenges exist for institutional investors to extract information that is likely to affect the stock price. The number of pages ranges from more than a dozen to more than a hundred. In addition, since they are issued before a shareholder’s meetings, they are issued in large numbers in a particular month, i.e., thousands of company notices are issued in June, the most concentrated month. This is a significant burden for institutional investors.

    The purpose of our research is to automatically extract pages that are likely to affect the stock price from shareholder convocation notices. To this end, we need to tag the pages to automatically extract what information is described on a page-by-page basis. In our research, we propose the following framework: We automatically create training data by a rule-based method and train the deep learning model that extracts important pages. We confirm the effectiveness of our framework for pages that cannot be extracted by the rule-based method.

  • 齊藤 史明, 山本 祐輔
    原稿種別: 特集論文
    2021 年 36 巻 1 号 p. WI2-H_1-13
    発行日: 2021/01/01
    公開日: 2021/01/01
    ジャーナル フリー

    This paper proposes Weasel Finder, a system that highlights weasel sentences while browsing webpages. The term weasel sentence is defined in the context of this paper as a quotation with an unknown or unidentifiable source. Following this definition, the system automatically detects weasel sentences in browsed webpages. Then, we investigate how highlighting weasel sentences affects the search behaviors and decision making of the users searching for information on the web. An online user study yielded the following results: (1) Highlighting the weasel sentences encouraged participants to invest more time in web browsing and to view a larger number of webpages. (2) The effect of (1) was more significant when participants were familiar with the search topics. (3) Web browsing elicited less change in the confidence of the search answers when participants were familiar with the given topics. The findings provide insights into how users can avoid gathering misleading on the web.

  • 坂井 明日香, 丸橋 弘明, 羽室 行信, 笹嶋 宗彦, 加藤 直樹, 宇野 毅明
    原稿種別: 特集論文
    2021 年 36 巻 1 号 p. WI2-I_1-12
    発行日: 2021/01/01
    公開日: 2021/01/01
    ジャーナル フリー

    Recently, data-driven sales management is widely recognized and sales at the real super-market is not the exception. For designing such strategies, first of all, we have to analyze consumers’ behavior. However, such an analysis is difficult, especially for the managers of the real shops, since they only have customers’ data of their own shops. Generally, the customers buy things not only from the managers’ shops but also other shops. The goal of this research is to develop a general method to transfer sales promotion strategy, derived from analysis on wide area, to local real shop. The authors analyzed such consumers’ characteristics who buy olive oils in Kansai region. For the analysis, we used QPR(Quick Purchase Report system, developed and managed by MACROMILL, Inc). Firstly, we divided the consumers on the QPR into five clusters, according to the simultaneous buying pattern. Then, we analyzed each of the clusters and found some emerging patterns of the purchasing behavior. Observing the patterns, we designed a marketing strategy for the real shop in Hyogo prefecture belonging Kansai district. Finally, we carried out an experiment at the shop to evaluate whether the strategy promotes the sales of the olive oil or not for six weeks. The result of the experiment showed that the marketing strategy is effective in one view. At the same time, we learned many lessons from the research, especially difficulty of the evaluation at the real shop.

feedback
Top