The readability of a paper document is often due to the “easy to handle” of the paper. In this paper, we analyze the action in ten kinds of readings on paper documents that occur in business (such as paper moving, page-turning). As a result, it was confirmed that various actions were performed at various frequencies for each reading. Moreover, in order to examine the reading support method from the viewpoint of the reading actions performed, we used cluster analysis of ten kinds of readings and obtained five clusters. Considering the support method for each of these clusters, the refreshed speed of the screen must be necessary for reading that frequently flips pages or moves between pages. Moreover, the lightness and thinness of reading devices must be remarkable for reading to move documents or to compare multiple documents. From our suggestions, it was clarified that the appropriate reading devices depend on the types of reading and that the supported functions depend on the types of reading.

Data analysis has been conducted in business operations in many fields. Analysis of text data like comments of questionnaires and reviews on the Web is also required. Most of the data analysis is conducted by using the software. The data analysts need to acquire knowledge of the software usage and mining operations. Some of the software provide instructions and tutorials about how to use. However, if the analysts are beginners of data analysis, the instructions and tutorials can not tell how to choose appropriate tools and how to operate the tools for their analysis purpose.

This paper proposes a presentation system of operating histories including new tools and operations to support for beginners of text mining. We use TETDM as the text mining software in this paper. We experimented the presentation system to verify the effect of the system on tool usage. The experimental results showed that the system could support the participants to use new tools to obtain analysis results.

This paper introduces a smartphone-based strolling support system that encourages photography during a stroll in a city. The system determines high-frequency photography areas based on the positional information of pictures shared on the Web. When a user enters one of the determined areas, the system notifies the user by vibrating the smartphone. The system only displays the relative location of the area on the screen and it provides no detailed information such as maps and routes. This feature is designed to enhance users’ awareness of the surrounding environment. In user studies, eight participants examined their use while walking around the city. The user studies suggest that the information provided by the system attracts users’ attention to surrounding environments and increases the amount of photography during their walk, especially for the users who do not take many pictures.

Hearing dogs can search for deaf people to alert important life sounds. The dogs use their own body and touch the people to inform them of the sound. Meanwhile, few of the dogs work actually. To solve the problem, a hearing-dog robot has been proposed and developed so far. Furuta et al. proposed a method to estimate a user’s life pattern. The life pattern is estimated using past searching data acquired from what the robot had searched for the user. The proposal aims to more quickly detect the user. But, lack and bias of the past searching data make the method weak. This paper proposes a method to estimate life pattern of the user. The life pattern is described as proper probabilistic distribution. To achieve the probabilistic distribution, the method applies a clustering algorithm for the past data. The method hires Dirichlet process mixture model for the clustering. We conducted a simulation experiment to evaluate the method. The experiment prepared a user model that has a certain life pattern on the experiment environment. We evaluated time cost for the robot to search for the user model. The experimental result showed efficient searching in comparison with the method of Furuta et al.

Since Lotfi A. Zadeh formulated the compositional rule of inference so called “extension and projection,” various kinds of fuzzy reasoning methods have been proposed with their mathematical properties. Mizumoto studied about many cases based on Max-⊗ composition. The cases lead to T-norm family of fuzzy reasoning methods. In the fuzzy modus ponens, ⊗ is used twice, when defining fuzzy implication and the process of composition. Under the normative criteria required for fuzzy modus ponens, we discuss which T-norms noted by ⊗ have good properties. In the usual modus ponens, when both “P implies Q” and the antecedent are valid, the consequent is right, but when the antecedent is false, we can say nothing about the truth of the conclusion. Also like the above case, unknown should be assigned to the conclusion inferred by fuzzy modus ponens, which is one of the five criteria.

The event-related potential P300 is one of the electroencephalography (EEG) features. Though the researches on the technologies using P300 have been widely carried out so far, there is a problem that the popularization of these technologies is difficult. Because “EEG devices” for measuring brain waves is expensive. One of the solutions to solve this problem is the use of a consumer EEG device which is not expensive. However, it is known that the measurement accuracy of a consumer EEG device is low. Then, in this paper, we carried out the verification on the measurement of P300 using a consumer EEG device and investigation of the P300 measurement for higher performance.

Linguists transcribe minor languages usually relying on their ears. This manual work requires time-consuming efforts and elaborate phonetic notation. In this study, we aim to construct a support system for fieldwork linguists documenting minor or low resource languages. In the previous preliminary experiments, we investigated sound matching between two languages using MFCC, but the correspondence accuracy was below expectations due in part to a lack of preprocessing data. Therefore, in this paper, we have worked on audio preprocessing using Wavelet analysis and SIFT features.

The purpose of this study is to enable machines to cluster images like humans do. We assume that human image clustering is based on the co-occurrence of objects. In this paper, we propose a scene clustering method based on LDA using the result of object recognition. Through the experimental result, we confirmed whether this clustering method is similar to scene clustering by human, or not.

Among various representative value indicators, mid-range (MR) has the disadvantage of being vulnerable to outliers, and is rarely used today. The MR index is one of the promising models used by humans to intuitively find the representative value of multiple observation values.

However, it is not efficient to use only the maximum and minimum values, so there is a possibility that the MR index can be an inappropriate model for small size data set.

In this study, we propose to treat a data set as nested range values and then calculate a weighted averaging value as new representative value indexes by combining the ranges with the set of weights created according to the definitions.

Among the four proposed models, the special extended mid-range index, which is named XMR, was selected as a model that does not require sorting of observation data. Then, its relative properties compared with traditional representative indexes (mid-range, mean, median, Hodges–Lehmann estimators) were investigated by a simple Monte Carlo simulation using exponential random numbers. Furthermore, preliminary psychological experiment was conducted, which resembled the situation of the simulation. The subjects were asked to reply their estimated representative value by intuition after reading 10 numbers which are assumed to note the intervals between severe accidents or disasters. The result of the experiment suggested that the differences among subjects’ estimation could be fairly large.

We aim to develop a system to detect Japanese learner’s mistakes automatically. In this paper, we proposed a method to detect mistakes by machine learning. Especially, we focus on mistakes that are grammatical errors which can be detected from an alone clause. We discussed features for machine learning and a learning method. Consequently, we found that random forest and features constructed by result of morphological analysis bring the best result, correct judgement rate is 76% for unknown data.

It is an important issue to develop an effective guidance system in a disaster. Several guidance systems have been proposed in addition to the facilitating emergency exit signs and implemented in some actual facilities using light stimulus or sound stimulus. Considering emergency situations that are raising in disasters such as blackouts, outbreak of smoke, and facility of damages, it is better way to implement several kinds of guidance systems using optical and sound signals for sight and hearing. The goal of our research project is to develop and implement an effective evacuation guidance system in which an egress root is instructed to evacuees using a sequential sound system. In this paper we discuss the recognition performance for direction detection of sequential sound emitting by the guidance system in order to discuss the effectiveness of the guidance system. For this aim the several experiments were conducted, in which the subjects judge direction of the sequential sound from the array of speakers under several conditions with respect to kind of sound, speed of sound, length between speakers and so on. The results of these experiments lead some characteristics for the recognition of direction of sequential sound.

In this paper, we propose an arousal-based sympathy expression model for educational support robots which provide learner-robot one-to-one learning to empathize with the learner. In the conventional sympathy expression model, when the learner gives a correct (or incorrect) answer continuously, the same emotion is repeatedly expressed, that is the emotion expression becomes uniform. In the proposed model, various emotions are expressed by giving priority to arousal in emotion selection. In the experiment, we compare a robot that expresses emotions at random, a robot with the conventional model, and a robot with the proposed model.

In this paper, we will consider how to configure AI which behaves differently from the original purpose of the game, namely, Out of purpose behavior AI (OAI). In the design of OAI, based on the idea that human setting goal of action on their own sense of values, we systematized “sense of values” to automatically adjust the goal setting according to the game’s play situation and the “attachment” that is the basis for the adjustment. Experimental results showed that OAI changed the play tendency to draw more attached shapes in the out of purpose behavior of drawing shapes on the map of maze games.

Onomatopoeias can simply describe sounds or state of things. There is a hypothesis that onomatopoeias have a characteristic that is called sound symbolism. Sound symbolism can make people imagine a specific image. According to this, similar specific sound and phoneme can make people imagine similar specific image. Urata et al. hired the sound symbolism and proposed an onomatopoeia thesaurus map, which can visualize semantic relationship among onomatopoeias. The onomatopoeia thesaurus map is constructed on a middle layer of an deep autoencoder. Urata et al. reported that the local positional relation on the map can visualize and indicate semantic relationship among onomatopoeias. But the evaluation about the visualization wasn’t satisfyingly performed. Our study set up a hypothesis based on Japanese linguistic knowledge about the sound symbolism, and experimented to evaluate visualization ability of the map. The most of the experimental result supported the hypothesis, however a part of it didn’t.

When using machine learning, it is necessary to prepare training data and set various parameters. In recent years, machine learning has attracted attention, and the number of cases where a person who does not have specialized knowledge or experience does machine learning has increased. SegNet, which is the target of this research, needs training images that annotated for recognition targets. Therefore, preparing training images requires an enormous amount of time and effort. Previous studies have shown data sets and parameters for learning each recognition target. However, there is no case where the investigation about how to give training images such as the number of training images and the setting of parameters required when using SegNet for the first time was conducted. The larger the number of training images, the higher the recognition accuracy can be expected, but the recognition accuracy does not necessarily increase in proportion to the number of the prepared training images. In this paper, we report the effects of the number of training images and the setting values of batch size on recognition accuracy as a way of giving training images.

We propose a method of generate tourist route for tourism support. The generate route defines a route as the order of visiting multiple sightseeing spots, and selects a route to visit the sightseeing spots within the specified time and shows the candidate routes. The location information and genre of each tourist spot are registered in the database. Location information is used to select a route, and genre is used to determine a candidate route. First, the selection of a route is done by creating a set of tourist attractions that can be visited within a specified time and determining the order in which the spots are visited. Next, focusing on the genres within each selected route, the routes are ranked so that if there are many different genres within the route, the ranking will be higher. Finally, the candidate routes with a specified number of routes are shown from the top of the rankings. We believe that this will help to revitalize tourism.

This paper proposes a method of automatically collecting a dataset for training a tweet sentiment analysis model. The proposed method automatically collects the training dataset based on the existence and the types of emoji. In this paper, we make a comparison between the sentiment analysis model trained with the automatically collected training dataset and the model trained with a training datasetthat has manually identified sentiment labels. The evaluation result demonstrates that, in terms of the performance of the trained tweet sentiment analysis model, automatically collected 9,000–136,212 tweets correspond to 270–540 tweets with manually identified sentiment labels.