We consider the task of variable selection in linear regression models. Because of their simplicity, linear regression models are widely used for prediction and forecasting. When applying linear regression models, it is important to conduct variable selection, for which we simultaneously select a subset of relevant input variables and optimize model parameters. By applying the SICM (Sequential Information Criterion Minimization) algorithm, which was proposed in our previous work, we propose a solution of the task. The algorithm enables us to continuously minimize an information criterion, which includes L0 norm such as the number of parameters, and was applied to logistic regression models and their mixtures. In this paper, we derive a method for continuously minimizing an information criterion in linear regression models.

In this paper, we propose to introduce a new input variable by using natural language processing for company information related to stock prices such as net news, and apply it to the prediction model together with the "open price", "close price", "high price", and "low price" of the stock price. As a prediction method, a multiple regression model and a neural network are used. Aiming for highly accurate stock price forecasting by applying the proposal method to the stock price forecast of multiple individual company stocks and comparing and verifying the effectiveness of the proposal method by simulation.

To greatly explode range of thinking for AI, it needs to be able to autonomously discover new alignment structures that can serve as a framework for inductive reasoning. A study aimed at a method for automatically constructing the three types of relations that support alignment structures was conducted. It was assumed that the specification relationship and their equivalence could be configured in various ways by computational procedures. However, the comparability could only be obtained from the properties of the sensors. Therefore, a method to obtain a new alignment structure by combining various specification relations that match the comparability of sensors was discussed. It can be regarded as "objectification" in the sense that it treats the object as an object that is easy to recognize and operate. And among the deep generative models currently presented, the model with the attention function looks satisfying the realization requirement for learning the object based on the alignment structure. However, learning of objectification itself still seems to be difficult.

In recent years, the Bayesian brain hypothesis has been proposed in the field of neuroscience, which states that Bayesian updating is used in decision making based on uncertain input stimuli. In behavioral economics, the anchoring effect has been known as a phenomenon in which judgments and estimates are influenced by numbers presented in advance, but Turner et al. showed that the phenomenon can be explained by Bayesian updating. Ozawa et al. found that the Turner et al. model assuming a normal distribution had problems when the amount of knowledge was small, and proposed a model combining logarithmic preprocessing and a t-distribution model extending the normal distribution to solve the problems. The result of Ozawa et al. is a model limited to size estimation that takes positive real numbers, but proportion estimation is also well known as a target where the anchoring effect occurs. Therefore, we hypothesize that the general anchoring model can be explained by changing the preprocessing appropriately and examine whether the model with the logit function as the preprocessing in the case of proportion estimation is consistent with the experimental results.

Analogy is an important cognitive process that supports human intellectual activity. It is useful in a wide range of mental activities that involve finding meanings in images, for example, mental activities such as metaphor comprehension, creativity, abduction, and transfer learning. There is the theory of indeterminate natural transformation (TINT; Fuyama, Saigo, & Takahashi, 2020) as a model of metaphor comprehension. TINT defines the meaning of a certain image as the sum total of associative relationships to the other images. Supposing that comprehension of a certain metaphor is a kind of creation of meaning, the process of metaphor comprehension can be expressed as the change of the associative relationships around and between the source and target of the metaphor. TINT requires a network of images connected by associative relationships. As a result of its computation, TINT rewrites the association network. So far, TINT has been implemented in a minimal form, and simulated and experimentally verified for only one metaphor comprehension process (Ikeda et al. 2021). Other than that, on the process of creating meanings, little research has been done about the time development of association networks. In this study, we analyze how the nature of the association network develops as TINT rewrites the associative relationship in the metaphor comprehension process.

In order to elucidate human perceptual functions, it is necessary to consider both bottom-up information processing, in which stimulus information received from the sensory organs is encoded into symbolic information, and top-down information processing, which is objective-oriented and based on memory, beliefs, and context. In this paper, we take the ResNet50 image classification problem as an example task, and conduct a basic study on the information processing when humans make judgments about visual information with ambiguities, and discuss the computation by which working memory during task execution penetrates the discrimination results of the model in a top-down manner.

Reservoir computing (RC) using asynchronously tuned elementary cellular automata (AT_ECA) has been shown to have high learning ability. The goal of this study is to propose an index for analyzing the reservoir state and to investigate the relationship between the criticality of AT_ECA and the learning ability by the index. The results of this study reveal that the high learning ability of AT_ECA based RC derives from its universal criticality.

We rapidly approximate a non-negative tensor with a rank-1 tensor. Although there have been many studies on rank-1 approximation, no algorithm guarantees that the resulting rank-1 tensor is the best approximation of the input tensor in the sense of the Frobenius norm. We find that any rank-1 tensor can be represented as a product of independent distributions when the tensor is viewed as a probability distribution. This property leads to a convex optimization formulation of the rank-1 approximation of a non-negative tensor, where we minimize the KL divergence instead of the Frobenius norm from input to the output tensor by projection onto a subspace consisting of products of independent distributions. Furthermore, we obtain an analytical representation of the best rank-1 tensor in our formulation using the property that some parameters representing a tensor do not change during this projection, which makes rank-1 approximation faster. The projection onto the space of products of independent distributions is widely known as a mean-field approximation, and our approach of rank-1 tensor approximation can also be viewed as the mean-field approximation.

In this study, we provides a convergence rate of a continuous black-box optimization algorithm, the (1+1)- Evolution Strategy (ES), on a general convex quadratic function, where convergence rate is decrease rate of the distance to the optimal point in each iteration. We show an upper bound of the convergence rate is described with the ratio of the smallest eigenvalue of the Hessian matrix to the sum of all eigenvalues. As long as the authors know, this is the first study which suggests the convergence rate of the (1+1)-ES on a general convex quadratic function is affected not only by the condition number of the Hessian, but also the distribution of the eigenvalues. Furthermore, we show a lower bound of the convergence rate on the same function class is described with the inverse of the dimension of the search space, which agrees with previous studies on a part of convex quadratic function.

Generalized Mutual Assignment Problem (GMAP) is a multi-agent based distributed optimization where the agents try to maximize the sum of the entire profits in the system. In order to cope with the difficulty to solve this problem within the reasonable computation time in a decentralized manner, a distributed heuristic algorithm to obtain feasible solutions is investigated in this paper. First, GMAP is reformulated by using the Lagrangian decomposition technique, then a multi-agent consensus based optimization algorithm is applied. Since it is too hard to obtain feasible solutions by the ordinary diminishing step size in the algorithm, an asynchronous and dynamic step size is proposed for the algorithm. Our numerical experiments show the effectiveness of our proposed method.

Machine learning-based modeling of physics phenomena is expected to accelerate simulations and to find a new phenomenon. Physics phenomena are often associated with conservation and dissipation laws of certain quantities. A dependable simulation must guarantee the laws of physics in discrete time. In this paper, we propose a deep learning-based modeling that ensures such laws of physics, and automatic discrete differentiation algorithm, which is an algorithm that ensures the laws in discrete-time. Experimental results demonstrate that the proposed framework ensures the energy conservation and dissipation laws up to the rounding error, and it learns a given dynamics more accurately than existing methods based on ordinary numerical integrators.

Scale-free networks are constructed by nodes and by links that represent places and connections respectively. Autonomous network exploration is essential to understand the structure of scale-free networks. The random walk model and the self-avoiding walk model are well known as representative network exploration models. The problem of the simple random walk model is that the agent revisits nodes repeatedly. The problem of the self-avoiding walk model is that the agent fails to revisit hub nodes. To solve these two problems, we propose the self-autonomous walk model as a new network exploration model. The self-autonomous model can return to hub nodes and is unaffected by network clusters. In this paper, we use the proposed model to find the average path length on the scale-free network and aim to improve the exploration efficiency compared to conventional models. As a result, the proposed model achieves shorter path lengths than conventional models.

Even when solving unknown tasks, humans are able to efficiently solve a wide variety of tasks in front of them by utilizing their knowledge and past experience they have gained in other domains. On the other hand, agents in Reinforcement Learning, which learns strategies based on rewards in an unknown environment, requires a lot of trial and error because it does not have knowledge of other environment and therefore cannot efficiently search using its experience. The solution to this problem is transfer learning, which is the adaptation of knowledge learned in another domain to a new domain. In this study, we focus on the cognitive function of analogy as a form of transfer. One model of analogy is the theory of indeterminate natural transformation (TINT) proposed by Fuyama and Saigo. It is an algorithm that constructs an appropriate functor by searching for natural transformations that displace the obvious functors in category theory. By using TINT in reinforcement learning, we aim to find a correspondence (functor) between the experience learned in another task and the experience in a new task.

Answer Set Programming (ASP) is an approach to declarative problem solving, combining a rich yet simple modeling language with high performance solving capacities. We here develop an ASP-based approach to Multi-Objective Vehicle Equipment Specification Problem (MO-VESP). The resulting system reads a MO-VESP instance of OVM format and converts it into a set of ASP facts. In turn, these facts are combined with a first-order encoding for MO-VESP solving, which can subsequently be solved by the ASP solver asprin. In our experiments, we succeeded in enumerating all Pareto optimal solutions for a small-scale problem.

In recent years, there have been many occurrences of earthquakes and other large-scale disasters. Since the Great East Japan Earthquake revealed that many residents do not know where to evacuate to in the event of a large-scale disaster, evacuation shelter allocation is currently an important issue. Okada et al. proposed a method for allocating evacuation centers using data on people's stay histories. However, this method has a problem that it does not take into account evacuation routes. In this study, we propose a method to determine whether or not the network in the area to be evacuated can be divided into buildings and evacuation shelters, and to obtain the allocation of evacuation shelters without intersecting evacuation routes. As a result of evaluation experiments, we found that the proposed method was able to avoid the crossing of evacuation routes, while the conventional method had the possibility of crossing evacuation routes.

Maximum probability partitioning is a classical problem of string, and the efficiency of the program to obtain the partitioning is important. This paper describes an efficient method for this problem and examine its efficiency by actual code. The efficiency is achieved by examining the behavior of frequency counting using suffix array.

The goal of this study is to visualize the relationships between classes and specialty (i.e., laboratory) for supporting students to determine the classes with the future direction of study. The students choose their classes by themselves because the university curriculum is highly flexible. Each class should be connected to some specialties though, it is difficult for students to understand the relationships from just syllabus without sufficient knowledge.This paper proposes a method to estimate the relationships between classes and laboratories in the faculty. The proposed method applies semi-supervised non-negative matrix factorization to reveal the common factors of knowledge in each combination of laboratory and class. It was suggested that reasonable results for the relationship between the laboratory and classes were calculated by the proposed method. We believe that it is possible for students to understand which class should relate to which laboratory

With the recent explosive increase in academic papers, researchers effectively understand the previous research by searching the highly cited papers in the related fields. On the other hand, new findings that overturn the concept of a field can be discovered through research that re-evaluates past research, which is different from the trend of the field. In this study, we propose a method to retrieve Sleeping Beauty, which is the paper that became cited after a long sleep, and Prince, which is the paper that triggered the discovery from Scopus whole dataset. The proposed method can exhaustively retrieve Sleeping Beauty and Prince pairs and find that Storyteller papers that cite Sleeping Beauty and Prince simultaneously propagate discontinuous discoveries to the field. Also, we found that discontinuous discoveries were more likely to be due to interdisciplinary research than to random citation though the rarity of the combination between categories had little impact.

When people are faced with a number of action choices, the issue of how to find an appropriate action is important to enjoy a comfortable life. In previous works, some models have been developed to predict social welfare, i.e., the similarity between a user’s interest and her action choice, by which the best action can be suggested for each user. Although they consider a scenario where a user’s interest may be affected by actions which she did in the past, they make a restrictive assumption that her interest should increase or decrease monotonically when similar actions are repeated. In this paper, we propose a new model that can mimic the scenario of boredom that user’s interest initially increases and progressively decreases through repeated similar actions. We fit the model parameter based on a real-world data, and discuss about the obtained result.

Pairwise comparisons have been widely used to estimate people's preferences. However, the preferences expressed by pairwise comparisons are not always accurate. Human evaluations are known to be distorted by cognitive biases. In the case of pairwise comparisons, the order in which the pairwise comparisons are conducted may have some effects on choice. We first confirmed the influence of the order of evaluation in pairwise comparisons by conducting a preliminary experiment. Based on the results, we propose a model that explicitly considers the effect of the order of evaluation to reduce its effect and to estimate the true preference. We applied the proposed method to the data collected via a crowdsourcing service and confirmed that the proposed model achieves better prediction accuracy scores than the baselines, especially when the number of items is large and the number of training samples is small.

The study proposes recommender systems which learn the distributed representation of items, sessions, and users from session data by using Item2Vec. For a recommendation query, the systems construct a session-specific distributed representation for the user (real-time user representation) in real-time via a simple computation method. In addition, we propose NN(Nearest Neighbors)-type and CF(Collaborative Filtering)-type search approaches which consider real-time user representations only and similar user representations, respectively. The experimental results suggest that the proposed systems are well balanced in accuracy, diversity, and novelty compared with the baseline systems. Moreover, CF-type search is superior to NN-type search in terms of accuracy, diversity, and novelty.

Recently, the market for consumer-to-consumer (CtoC) transactions has been growing and attracting attention. To revitalize the CtoC market, existing studies have mainly focused on predicting the departure of sellers and analyzing the purchase intention of buyers, and little research has been conducted on the trend of sellers' revitalization. The purpose of this study is to investigate the effect of the relationship between sellers on the change in seller's sales. We created a network with sellers as nodes, extracted communities, and analyzed the relationship between changes in sales and features of sellers for each community. We found that the impact of popular sellers on neighboring users varies from community to community, in some cases promoting activation and in others inhibiting it.

By visualizing the influence of an individual in a group and using it to recognize one’s own influence, it is possible to improve one’s sense of ownership and belonging to the group. This will contribute the enrichment of human-relationships and the proper growth of the group. In this study, we took into account the characteristics of SNS and calculated one’s influence in a group based on interactions on Twitter.

A crowdsourcing with cooperative collaboration among experts should be performed by teams considered with the compatibility of crowd workers. In this study, we develop a system to automatically form an organization which can efficiently carry out complex and large-scale projects in crowdsourcing. The system represents the compatibilities among workers based on their past collaborative works as a social network. We also develop an algorithm based on the greedy method to search teams with the compatibility optimization under the constraints of budget and skills. In the experiment, we collect 169627 users on 33983 repositories from GitHub, and determine the social network and skills of the workers by the used programming languages and contributions for the repositories. The characteristics of the developed system are observed by the simulation experiments with a virtual project with skills and budget requirements. In particular, as a result of comparison on teams formed with the minimum budget, the proposed team formation algorithm achieves 96% higher in the compatibility of teams compared to the algorithm without considering compatibility.

Crowdsourcing has been widely used to efficiently obtain labeled datasets for supervised learning from large amounts of human resources at low cost. However, to obtain high-quality results through crowdsourcing, the challenge is to deal with variations and biases caused by the fact that the work is performed by humans. We focus especially on the bias of workers in selecting tasks to work on. Workers may be biased in choosing tasks to work on based on their circumstances and preferences, and this may affect the quality of results. We propose to reduce this bias by using the observation bias reduction method, which is used in causal inference, to improve the accuracy of results by aggregating responses. Through experiments using artificial and semi-artificial data, we verify the existence of bias and confirm that the proposed method improves accuracy under certain conditions.

The remarkable development of AI technology has made it possible to support a variety of real-world decision-making problems. However, the power of AI is still limited for open questions with ambiguous definitions and with a limited amount of data. On the other hand, the method of solving problems through collective intelligence by calling for cooperation from many people via the Internet has been widely used in recent years, but this method has reliability problems. In order to increase the reliability of decisions made by increasingly complex AI systems, much research has been done to make AI decisions interpretable to humans; similarly, the interpretability problem of seeking a basis for judgment by unspecified collective intelligence arises here. In this research, we aim to solve this problem by using Analytic Hierarchy Process (AHP), which is used as a method to aid rational decision making, as a framework for collecting and evaluating answers and criteria using crowdsourcing.

Infection with COVID-19 has been confirmed in almost all countries and regions including Japan. Under these circumstances, it is important to measure people’s awareness and willingness to go out. In this study, we analyzed the changes on Twitter from various perspectives. As a result, we found that people’s interest gradually diminished from the first wave to the third wave, as shown by the decrease in the frequency of tweets and the amount of retweets. In addition, there were qualitative content differences in the tweet tendency when comparing with the period of infection.

The open access of research papers has become widespread, and opportunities for non-researchers to access the papers are increasing. During the COVID-19 pandemic, papers of the author's version are sometimes published on preprint servers, and there are many tweets about these preprints on Twitter. In this study, about the mention to preprint servers on Twitter, we focus on three sites, arXiv, bioRxiv, and medRxiv, and conduct preliminary analysis. As a result of the analysis, we found that many mentions of arXiv were not retweets and that many bots were active. We also found that medRxiv was mentioned by a large number of users, and it tended to be more noticeable than other preprint servers.

While SNS has been widely used in recent years, there are offensive posts such as slander are made, which has become a problem. Therefore, it is necessary to devise effective countermeasures after understanding the characteristics of aggressive posts such as slander. The purpose of this study is to build a model that can accurately express the dynamics of occurrence and convergence of aggressive posts such as slander on themes that divide opinions, and analyze their characteristics. The model defines five states for each user: neutral, agree, agree and aggressive, disagree, and disagree and aggressive, and is set to take into account the influence of neighboring users on state transitions. As a result of measuring the time-series distance between the development of user who is in the aggressive state on Twitter and the development of user who is in the aggressive state in the model by DTW, the proposed model is 4.493 for agree and aggressive, 7.443 for disagree and aggressive. Therefore, it was possible to construct a model with a high degree of similarity compared to 132.2 and 170.0 when each state was randomly transitioned. It was also shown that the cumulative ratio of the total number of users in aggressive states decreases when the influence of users that aggressive and have same opinion is suppressed.

While the rapid spread of social media has greatly advanced social interaction, aggressive behavior, such as abuse, has become a problem. In some cases, abuse can lead to mental illness or suicide, and it is important to understand the mechanisms of aggressive behavior and detect them in society. Text-based analysis of attacks on social media has become difficult to characterize due to the shortness and ambiguity of the sentences. In order to solve this problem, user-based analysis of attacks has been conducted, but the results are diverse and no certain knowledge has been obtained. In this study, we analyzed the attack tweets extracted from the sampling data of Japanese Twitter using sentence-BERT in order to classify the attacks into several types based on the nature of the users and their network relationships, and to clarify their characteristics. We found that most of the attacks occurred when the users' networks were close to each other, and that the users who attacked distant users on their networks tended to post many negative tweets. We also found that the users who attacked distant users tended to post more negative messages.

Virtual YouTuber (VTuber), the virtual characters of video contributors, are becoming more and more common. As of November 2020, there are more than 13,000 VTubers, and their activities are attracting attention from many fields talents as celebrities and creators. However, due to their various characters and their diverse themes of the videos they post, it is difficult to grasp the present situation of the activities of VTubers. In this paper, we propose a VTuber search support system. This system consists of a search function for VTubers and a page displaying information about VTubers. The search function allows you to search for VTubers by either their names or keywords that describe their activities. On the VTuber page, you can search for VTubers by using related keywords and VTuber information. This system has made it possible to actively explore VTubers and deepen our understanding of them.

It is called meta-inference that player infers based on information that cannot be known only in werewolf game. For example, inference using the player's skill of werewolf game is called "person meta inference". In AIWolf competition held by AIWolf project, one set consists of 100 werewolf games with fixed members, so it is possible for agents to make meta-inference within the set. Therefore, in recent years, many of the top winning team of the competiton adopt meta-inference. In this paper, utterance pattern feature is introduced in order to estimate the meta information of the agent's team. The effectiveness of this feature is shown by the significant improvement in accuracy of the DNN which has this feature as input. It is also reported that the agent which uses this feature achieved good results in latest international AIWolf competition.

Basketball players generally have fallen into one of five positions. Recently, however, the roles of those positions have become ambiguous and there exist players with different playing styles even in the same position. In this study, we aim to classify the playing styles of 6,600 players who have played in the past 20 seasons from 2000-01 to 2019-20 by visualizing their playing data with Mapper networks. In addition, we compare the differences of playing styles in the the past 20 seasons by quantifying the results of the visualizations which are represented by networks. As a result of the analysis, we classified NBA players into 11 playing styles. Moreover, we found that the differences in playing styles among players have increased over time.

These days, many novel neural networks for modeling spatio-temporal relationship are proposed as many kinds of spatio-temporal datasets like location dataset or traffic dataset are published and utilized. However, novel networks have a common problem that they cannot handle properly multimodal data with complex (multi-step) relationship, e.g. Modal A affects modal B and modal B affects modal C. This problem must be solved because much more kinds of spatio-temporal data will be distributed in the future. In this paper, for discussing what kind of structure “multimodal spatio-temporal network” should be, we conduct some preliminary experiments which includes extending existing spatio-temporal network to handle multimodal data and comparing prediction capability with the original network. Based on the results, we conclude that “multimodal spatio-temporal network” should properly encode the information which affects relationship of modals dynamically, e.g. meteorological data.

Toulmin Model is known as a representation of the logical structure. The one composed of the three elements Data, Warrant, and Claim is called Triangle Logic Model. Triangle Logic assembly activities system using this model has been developed and it has been suggested to promote logical thinking. However, since it haven't done any activities to examine the validity of the assembled logic, it is hard to say that it also supports critical thinking. In this research, we aim to support critical thinking and to incorporate formal fallacy by triangle logic into triangle logic. We also aim to identify formal fallacy as the proposition used in reasoning in triangle logic matches the formal fallacy rule. As a learning task, after assembling triangle logic as a reconstruction task of the logical structure of another person, it is made to judge whether the triangle logic belongs to formal fallacy, and if it is judged to be formal fallacy, which one you will be asked to determine which formal fallacy rule the proposition applies to. This task will allow learners to examine the validity of the logical structure of others while looking back on their own logical structure, and can be expected to promote critical thinking.

In the treatment of infectious diseases, rapid and correct identification of the inflammatory bacteria is important for the Antimicrobial Stewardship. In the numerical identification method, multiple biochemical reactions of the inflammatory bacteria are tested, and the positive/negative results are accumulated. The metabolic profiles are compared with established database obtained by testing known strains, and the species is estimated by calculating likelihood. In many clinical laboratories, automated bacterial identification devices based on biochemical reactions are used. RAISUS S4, developed by us, is one of them. It monitors time series data of biochemical reaction as fluorescence value and identifies bacterial species by numerical identification method. We conducted a study to establish a rapid and accurate testing method by using the fluorescence value more effectively. The accuracy of the neural network model using the fluorescence values as input and the bacterial species names as output were verified. The results showed a high rate of correct identification for both Gram-negative rods and Gram-positive cocci within 2 hours of incubation. This method follows the conventional method of classifying bacterial species based on biochemical reactions, and can also handle time series data, may be an effective way in clinical laboratories.

The purpose of this paper is to predict an increase or decrease in the number of event news to capture propagated information by using GNN. Also, we aim to realize different influence for each type of event by using embedding vectors of ”news tags” as inputs to the model. Moreover, GAT attention makes us to interpret the news impacts from other countries. Our experiments showed that GNN captured information propagated from news in other countries and improved the prediction accuracy. Furthermore, the level of attention was qualitatively consistent with some event samples.

In this study, we analyzed contents of TV news program in japan, in order to clarify relationship between COVID-19 infection status and how these news were covered. First, we clarified the characteristics of COVID-19 coverage by analyzing changes in the coverage time and viewership, changes in the content of coverage, and changes in the priority of COVID-19 coverage, etc. Then, we analyzed how contents of TV news program changed in response to the increase/decrease in the number of infected people on the first, second, and third wave of infection. As a result, it is cleared that TV media and people's interest was big changed after first wave.

We propose a method to extract areas potentially spreading COVID-19 by focusing on correlation between mobile phone population statistics and the number of new positive cases. Our experiment showed that our method can successfully extract areas that are consistent with the government's views on infection sources.

With the recent exponential increase of academic papers, the importance of indices for extracting academic paperwith high information value has been rising. For example, the total citation is used as a general index, and anindex that measures the innovativeness of academic paper from citation networks has been proposed. On the otherhand, academic paper with a high degree of cross-disciplinarity is attracting more attention due to its contributionto complex social issues. In this study, we propose an index for evaluating the innovativeness of academic paperthat takes into account the cross-disciplinary nature of citations by constructing a citation network from academicpaper data in Scopus and quantifying the distance between disciplines. From this index, we extracted the papersthat had an impact across disciplines, and also clarified that the number of people in a research organization canaffect the cross-disciplinary nature of the research.

The PICO (Population, Intervention, Comparison, and Outcome) framework is generally used to design clinical trials and researches. Since PICO information is not explicitly labeled in typical medical article texts, automatic identification of PICO information can improve efficiency of medical literature reviews. Although various PICO extraction models, including those based on BERT, have been reported so far, PICO information is sometimes overlapped with each other and none of the previously reported models could extract overlapped information simultaneously. Here we propose an alternative approach in which a layered LSTM model for nested named entity recognition is used in combination with BERT models pre-trained with medical domain texts. We show that the proposed model has comparable performance with the previously reported models, and can simultaneously extract overlapped PICO entities.

Synchronous movement strengthens a bond with others, makes positive impressions, and induces related behaviors. This study investigated how whole-body synchronous movement influenced impressions of exercise activity with a human-like agent and behaviors. In the experiment, participants synchronously or non-synchronously repeated sit-down and stand-up with the agent through the rhythm of metronome. It was implemented to recognize body movement and react using Kinect, based on the future work of our previous study. The post-questionnaire result showed that participants in the synchronous condition made a positive impression on using the agent in the exercise activity more significantly than those in the non-synchronous condition. We also discuss social application of a human-like agent as in this study.

One way to improve the relationship between humans and agents is to have humans sympathize with the agents. By empathizing, humans take positive and kind actions toward agents, and emphasizing makes it easier for humans to accept agents. In this study, we focused on self-disclosure from agents to humans in order to realize anthropomorphic agents that elicit empathy from humans. Then, we experimentally investigated the possibility that an agent's self-disclosure facilitates human empathy. We formulate two hypotheses and experimentally analyze and discuss the conditions in which humans have more empathy for agents. This experiment is a three-way mixed plan, and the factors are 12 conditions of appearance, self-disclosure, and empathy values before and after watching the video, 2 × 3 × 2. Participants were asked to complete a 12-item questionnaire after completing two tasks, and the promotion of empathy for human agents was analyzed. As a result, we found that the appearance factor did not have a main effect, and self-disclosure, which is highly relevant to the scenario used, facilitated more human empathy with statistically significant difference. We also found that no self-disclosure suppressed empathy. These results support our hypotheses.

In the era of digitalization, there has been growing interest to utilize real data involved in manufacturing. One challenging issue in manufacturing lies in standardization of human operations like assembly. The working time in them can vary, according to the operation complexity and operator skill. In the following, we call this variance "work deviation''. To quantify it, we address the elementary task to identify a series of assembly processes from video data. In this paper, we first prepare for one benchmark simply capturing normal assembly works. Using the benchmark, we empirically evaluate the proposed method based on pose estimation for identification task. We prepared query about the working time and compared the query results by the manual result and the machine learning model result.

Designers create representations aesthetically pleasing that achieve a functional meaning. Color is one of key elements that plays an important role in design, both functionally and aesthetically. Thus, for designing color schemes, it would be important to understand the information associated with the color in order to create an expression that has functional meaning. To address this issue, we collected data on the association of color with words and shapes from more than 10,000 Japanese speakers by a series of multiple-choice questionnaires. Through a multivariate analysis, we were able to represent the association of words, shapes, and colors across several different semantic categories as a two-dimensional map reflecting its topological structure of the map. Using this result, we propose a method to generate linguistic expressions for color combinations by vector summation in a word2vec vector space and discuss its utility for design.

Recently, the necessity of "uniform test forms" for which each form comprises a different set of items,but which still has equivalent measurement accuracy has been emerging. However, the construction of uniforms tests often suffers bias of item exposure frequency. This problem decreases reliability of items and tests. Therefore, the item exposure frequency should have a uniform distribution. For this purpose, this study proposes a new uniform test assembly: uniform test assembly with lower item exposure using integer programming. Our method considers bias of item-exposure by excluding highly exposed items from an item bank when generating a uniform test. We demonstrate the effectiveness of the proposed method using simulated and actual data.

Infographics convey information in attractive and effective way by transforming and combining various types of data, such as texts and pictures. Today infographics are often created using computer software. However, these tools generally require a user proficient skill and deep understanding upon operation. In this paper, we propose a user-friendly assistance system for creating infographics by using natural language as its input. Infographics created by this system contain numerical expressions and icons based on important numbers and nouns that included in the given texts. In addition, considering that the same information can bring different feelings to people by changing a writing style of texts (such as pessimistic/optimistic) and numerical expressions, our system provides a function to select a base color of infographics as "positive" or "negative". To evaluate our method, we conducted a questionnaire survey comparing infographics of different colors that assumed to affect an impression of information. As the results of analysis, it was found that the color-controlled infographics are much more efficient to enhance the impression in the intended direction than comparative images without color control.

In last few years, the Internet and Web contents have become remarkable tools for studying. However, most search engines that can find Web contents applicable for studying are not beginner friendly. Learners must manually compare several pages on the search engine to find beginner friendly Web contents. Visual intelligibility in Web page layout and beginner friendly Web page texts are the requirements of Web contents for beginners. In this paper, we develop a dataset of Web pages explaining academic concepts, to which we manually annotate their knowledge amount and learning level. This paper especially focuses on math and science academic fields such as statistics, calculus, linear algebra, mechanics, electromagnetics, chemistry, programming, and IT. In those academic fields, we collect major Web sites explaining academic concepts and manually annotate their knowledge amount and learning level to Web pages of those sites. Finally, we analyze the knowledge amount and learning level of each of those collected major Web sites.

The purpose of this study is to develop a quantitative measurement method for humans' concentration on learning using an image sensor.This paper investigated the correlation between the concentration obtained by EEG measurement and the body motion obtained by the image sensor. The brain workload calculated from the EEG measurement results was used to measure the degree of learning concentration. The auditory steady-state response (ASSR) was used to estimate the workload. Two Kinects as the image sensor were used for the body motion measurement, and one was placed in front of the subject (Master) and the other was on the left hand side (Sub). Eleven healthy Japanese people participated in the experiment. The correlation between the calculated workload and the body motion information was investigated for two learning tasks. As a result, we succeeded in statistically showing the body parts that have a huge relation to the degree of learning concentration. In the future, we aim to develop algorithms and implement software application that can quantify the degree of learning concentration using only one Kinect.

This paper describes work in the EU-Japan collaboration project e-VITA in the context of Active and Healthy Ageing. The goal of the project is to prototype Virtual coaches to sustain older adults’ well-being in smart environments, develop standards and norms for interoperability, and conduct a proof-of-concept study to assess user-acceptance in real-life environments. We will focus especially on the interaction aspects enabling the users to conduct dialogue interaction with the Virtual Coach and the integration of various system components into a federated AI platform which supports the development of the Virtual Coach.