We developed a prototype of trekking stock that can measure the pressure and acceleration during climbing gait. In both wearable devices (eyeglasses / wristwatch type) and prototypes, it was possible to measure walking data of climbers on actual mountaineering paths, and it was shown to monitor climbers' biological information and walking condition. Furthermore, as a result of the evaluation experiment, we could obtain new knowledge on "visualization of stock movement" and "difference in stock utility in climbing and descent".

In recent years, customer service robots have been rapidly deployed. However, it is still unclear what kind of behavior is effective for sale. A conversation cannot be initiated, if the customers are not interested in the robot. Therefore, in this paper, we focus on engagement in the field of Human-Robot Interaction (HRI). In order to produce socially intelligent responses, we introduce a machine learning method to calculate a degree of engagement by analyzing the data which we conducted a field experiment in a real shop. As a result, we develop a model that the robot can hence decide on a socially intelligent response in the start of the conversation.

Telepresence robot can give a sense such like remote person is really existing there. However, this sense is suddenly disappeared after the call. Therefore, the talker will feel loneliness. This research explores "tele-vestige" in which talkers can feel weak presence even after a call finish. To give this "tele-vestige", remote person's face is morphed to a telepresence robot's face after the call. We tested a pilot experiment to investigate the effect of this morphing on loneliness.

In Japanese, it is difficult to recognize completion point only by syntactic information because Japanese is the language whose case structure is not determined by word order. However, turn-taking occurs smoothly and there is little possibility that long silence or duplication occurs between utterance.I claim that non-verbal behavior such as non-speaker's gaze and backchannel is important to recognize completion point of turn-taking in multi-party conversation. In this paper, we observed an excerpt of dialog from the video recording s of the discussion. I found that there were differences in gaze of non-speakers between when speakers switched and when the speaker continued.

To make it behave as if an artifact (agent) has an internal model (model of others) is important as a design principle for making people understand the artifact (AI system) with increased autonomy. In particular, in this research, we focused on "time" and proposed a model that allows people to assume other models of artifacts by temporal overlapping of utterance behaviors in communication. In this paper, first of all, after overviewing our research on "duality of the body" and "temporal overlapping" in the interaction, we proposed a framework of "models of others emerging by temporal overlapping". We conducted experiments based on this framework and verified its validity. In the future, from this point of view, we are planning to conduct research on the interaction design of people and artifacts based on "time".

Recently, there are a lot of communication robots that are aimed to be a good human partner. We believe that these robots can also be a good tool for inducing the deep self-understanding throughout the conversation with these robots. We sometimes project ourselves to communication robots and this external self-projection helps to know about ourselves objectively. In this presentation, we would like to propose the model connecting external self-projection to behavioral alternation. Further, we also discuss how a robotic system can facilitate the behavioral alternation by introducing our several recent findings.

Story is a fundamental communication style in human being. Story comprehension is an immersed experience in the story world, and works as a simulation (or virtual experiences) in the real world. Thus, readers can learn how to understand story characters’ mental states. If it is true, literary reading should be a training of social abilities in the real world. We used "strange story task" as a near transfer measurement (verbal task), and "animation task" as a far transfer measurement (nonverbal task). The results suggest that literary reading enhances social abilities in the strange story task. For future, the story comprehension in the virtual reality or augmented reality would be promising to develop the pervasive training program for stoical abilities.

Discovering new materials that possess on-demand properties is the central demand in every industrial domain. We constructed the first full-stack material discovery system consisting of several technical pieces; feature encoding, regression, solution search, and structure generation. Those pieces are coordinated to coherently work by newly defining two kinds of feature vectors; data-driven feature and pre-defined feature, and developing an algorithm to generate molecular structures by using those feature vectors. The capability of the system to discovery new molecules is demonstrated by a public dataset of commercial drugs.

Analyzing communications between the honeybee waggle dancers and their followers in their hive is one of the most important and interesting issues to reveal a mechanism of honeybee’s language. In general, these behavior analyses have been usually conducted by extracting honeybee’s walking trajectories from recorded long-time video data manually. To decrease the hard work of observers and their artificial errors, we have previously proposed an automatic tracking algorithm of multiple honeybees using image processing. Besides, we have constructed an automatic recording system for long-term tracking of honeybee behaviors in an observation hive. Using this system, we recorded the hive and its corridor to the field over 4 weeks in June 2016. Finally, we obtained video data over 20TB per a month. Analyzing honeybee’s behavior from tahis enormous amount of data is required an extremely long time even if using a high spec computer. In order to deal with this issue, we first extracted the area and time of waggle dance from the recorded video data using a preprocessing based on frame-difference approach. Then we applied our tracking algorithm for the extracted partial video data. In the preliminary experiment, we conducted the automatic tracking of the waggle dancers and their followers for 13-hour video data.

This paper is concerned with a graph grammar that can be inferred from data and always generates valence-consistent molecular graphs. Our result is that the requirement above can be satisfied by a hyperedge replacement grammar inferred from molecular hypergraphs, which we call a molecular hypergraph grammar (MHG). By substituting MHG for SMILES (a grammar generating a string representation of a molecule) in a generative model of molecules, we can generate novel molecules without decoding errors, which have been one of the common issues when using SMILES.

Due to the improvement of the performance of computers, GA is actively applied to actual engineering problems. Most of engineering problems are constrained optimization problems. Penalty method is well-known as the solution to such constrained optimization problems. The benchmarks of constrained optimization problems have only small number of constraints. Thus, the effectiveness of penalty method has not been investigated in many constrained problems. In many constrained optimization problems, penalty method has a risk that all constraints cannot be satisfied. This paper proposes the stepwise satisfaction method of constraints to satisfy many constraints. In the proposed method, the priority of constraints to be satisfied is defined based on the initial population. Furthermore, this paper studies on dividing constraints into difficult ones and easy ones and combining the proposed method with the penalty method. In the experiment, the performance of the proposed method and the penalty method was compared in two problems.

In this study, a deep Radial Basis Function (RBF) Group Method of Data Handling (GMDH)-type neural network which has the deep neural network architecture, is applied to the medical image diagnosis of liver cancer. Deep RBF GMDH-type neural network has abilities of self-selecting the number of hidden layers, the number of neurons in hidden layers and useful input variables. This algorithm is applied to medical image recognition of liver cancer and it is shown that this algorithm is useful for medical image diagnosis of liver cancer and is very easy to apply practical complex problem because deep neural network architecture with many hidden layers, is automatically organized so as to minimize the prediction error criterion defined as Akaike’s Information Criterion (AIC) or Prediction Sum of Squares (PSS).

Feature Selection is an important preprocessing step for pattern recognition and data mining problems. This process selects necessary features and removes redundant features. In this work, we propose a new fitness function for feature subset evaluation. The proposed new fitness function is used for feature subset selection with EC algorithms. Simulation experiments using the benchmark datasets have been done and the results are compared with popular methods.

A three-dimensional (3D) entire shape reconstruction method, which performs simultaneous 3D registration of multiple depth images obtained from multiple viewpoints, is proposed in this paper. Unlike most other 3D registration methods,the proposed method fully utilizes a silhouette-based objective function taking out-of-view and non-overlapping regions between two models into account as well as depth differences at overlapping areas.With the combination of the above functions and evolutionary computation algorithms, the entire shape reconstruction from small number (two or three) of depth images, which do not involve enough overlapping regions for other 3D registration methods, can be realized. A CMA-ES algorithm with partial restart strategy to speed up the registration process is proposed in thispaper.

This paper discusses about the characteristics of the Neural Network with Cell Differentiation Algorithm by assuming the simplified model as a virtual software and examines its characteristics. Several conditions were found for making them larger and possible to acquire various recognition abilities.

In recent years, SNS has become widespread and it has become possible to exchange easily with various people through a network.In order to carry out smooth exchange it is necessary to understand the person ’s character, but only the sentence on SNS Since it is difficult to understand sufficiently, it is easier to understand the exchange partner by estimating the character of the author from the characteristics of the sentence.In this research, for a Twitter user, a set of sentences posted on Twitter to estimate personality by deep learning and to present the estimation results to users. We constructed a classication network by deep learning for each of a plurality of elements of human character and developed an estimation system.As a result of evaluating based on correct answer data created by cooperation of plural people, High conformity rate, recall rate, and sufficient validity could be conrmed.

This paper reports a study on creation of supplementary learning materials for foreign language by using comics. We propose a method to make a learning material for role words of Japanese and make an interface for learning the role words. We experimented with the interface for verifying the efficiency of the method. Participants of the experiment learned 30 role words for three days and took a test to answer appropriate role words. The participants with the interface marked higher scores than scores by participants who learned role words by using textbooks.

In recent years, with the increase of users of social media, many people are conducting online communication. However, recently troubles caused by inappropriate posting in social media are a problem. Considering these current situations, it is important to write sentences considering the emotions of the readers when conducting online communications in order to build good relationships.Therefore, in this research, using sentences as input, we implemented a function to visualize words expressing favor or malice of a writer and a function to urge text modification, with the purpose of supporting writing sentences that the writer receives a good impression from the reader. As a result of the experiment, it was found that the writing sentences support function of this system can support writing sentences receiving good impression from the reader.

This paper addresses the detection of overgeneralization of be-verb found in learner English. It is an error where the subject and complement are not semantically equivalent in a be-verb sentence as in *Paris is rain. This paper presents a method for detecting it by predicting through word embeddings whether a given subject and complement pair is semantically equivalent or not. This paper also presents a method for determining the hyperparameters in the method efficiently and effectively. Experiments show that the present method outperforms two baseline methods based on corpus statistics and WordNet ontology. Looking into the detection results brings out a way of generating feedback messages for learners that explain why the detected error is not a valid English expression.

In this research, we propose a method to extract rare cause-result expressions from summary of financial statements. For example, our method extracts effect expression that "Sales of drinking paper containers has increased" with respect to cause expression that "due to hot summer", as a rare cause-result expression. It is difficult to imagine that the sales of “drinking paper containers” may increase in the case of “hot summer”. Our method calculates the conditional probability with words contained in the result expression co-occurring with words contained in the cause expressions. Moreover, our method determines that cause information is contained in the result expression or not by using deep learning. Our method extracts rare cause-effect expressions by using the conditional probability and the cause information.

In this paper we study the capabilities of ccg2lambda to perform semantic parsing and natural language inferences in the financial domain. We observe that the system is conveniently modular and highly interpretable for humans since it includes graphical visualizations of semantic representations; these visualizations allow us to trace the process of semantic composition and easily identify points of failure. Moreover, it produces results in XML which makes it easy to integrate in larger, in-house systems. The system separates the compositional mechanism from the specification of semantic theories, allowing extensions to handle new linguistic phenomena with a relatively small effort. Finally, we present our approach to adapt the system to the specialized vocabulary that arises in the financial domain by using axioms and ontologies.

Coherence of the document is having unity of semantic, further It is necessary to write a document that is easy to understand. It is difficult to evaluate coherence of long document, since demand of automatic evaluation of coherence is high. Also, previous method focused on adjacent sentences to evaluate coherence. However, it is necessary to consider document structure, since document consists of multiple segments. Therefore, in this research, we propose metrics measuring coherence based on sentence similarity graph with document structure. We performed comparative experiments with previous method. As a result, some output accuracies comparable to previous method, it is assumed that proposed method is valid to evaluate coherence.

This paper defines textual evidence inference problem, a new NLP task for decision-making application. The objective of the textual evidence inference problem is to classify a given text into evidence categories based on a given decision-making question and a given set of sentences representing values and preferences of decision-makers. We report the technical difficulties of the problem by using newly developed dataset.

Authentic English passages are not always appropriate for learners due to their vocabulary level; hence teachers sometimes have to modify the text by making sentences simpler or replacing difficult words with easier ones. This process, however, takes time and could be a burden for teachers. The present study aims to build an automatic lexical simplification system that can assist teachers in preparing materials for classes and examinations. The proposed system first selects target words based on CEFR levels and then lists candidates from a thesaurus. Then, the paraphrasablity of each candidate is examined using a word embedding method. The results show that the proposed method can provide correct candidates for more cases than the baseline and existing methods and is robust even when the target is a polysemous word.

With the progress of the technology of artificial intelligence (AI), there have been various groups that discuss the problem of ethical, legal and social issues provided by the AI technology. Although they cooperate with each other to formulate guidelines for the development of AI products, the variety of these guidelines will require great effort to international companies to follow all of these guidelines. One possible factor of such differences is the national character, namely the aspect emphasized in these guidelines is different depending on the country or area in which they are discussed. To examine this, we show the difference of these guidelines and discuss it focusing on a matter of national character.

Although the importance of AI ethics has been increasingly recognized, it is hard to say that the results of the discussion is incorporated into research and development of AI technologies, so there is a gap between them. In this research, we provide an organic and dynamic AI ethics library with the aim of resolving this gap by supporting the practice of ethical design by AI engineers. Here, organic means that it deals with complex relations among different AI ethics discourses. And dynamic means that, in interaction with users, it dynamically adopts new issues and helps users think in their own contexts. For example, AI ethics library visualizes a structure of each AI ethics discourse in the standard form, presents the distance and relationships among the discourses, and seamlessly connects them on the extension of AI technologies. Furthermore, AI ethics library is expected that the AI ethics side can also be reconstructed as a more practical one through the practices. In this paper, with clarifying the framework of AI ethics by applying the ethical design theory, we introduce the overview and cases of the AI ethics library and evaluate its effects.

The purpose of this study is to extract new guideline items about AI development by analyzing ELSI (Ethics, Legal, Social Issues) focuses on service utilized AI from the view point of service provider and service user. Use case (sample service) in healthcare area is considered and interview related to this service is carried out. From the result of the interview and analysis, five guideline items modified from original AI development guideline and seven new guideline items are extracted. To the future, valid verification of these items will be carried out through extension of target area and evaluation.

This document reports the methodology and process of inter-disciplinary workshops involving various stakeholders to discuss IEEE Ethically Aligned Design (EAD) report. EAD report aims to move beyond the excessive fear and expectations associated with AI and increase innovation by creating ethically aligned designed AI. Authors organized the report items into 6 workshop series. The goal of our workshops is to understand the content of EAD and create a network among AI/IT researchers, social science and humanities researchers, industries, policy makers and other stakeholders.

In Japan, the revised Personal Information Protection Act is enacted in full force on May 30, 2016. It became possible to provide personal information to third parties under certain conditions such as anonymization without user agreements. However, in the case of no limitation on background knowledge of attackers, complete anonymization maintaining usability of data is impossible in principle. It is difficult to establish a rational anonymous processing standard. In this paper, an anonymization method based on residuals to establish both anonymity and usability is proposed. Moreover, it is shown that to apply anonymization with ambiguous purpose of use is meaningless.

In the near future, it is thought that it is necessary to coexist a robot and a human being. In order for humans and robots to coexist within the same society, smooth communication and interaction between the two sides becomes essential. To that end, it is an important task to develop a user interface design, that is, an interface with weights placed on the user side. In order to derive the design requirement of the robot, we attempted to derive the robot design items from the science fiction movie as a material, considering the human psychology of the robot opposite the robot. As a result, in addition to existing design items for determining design and specification, arrangement of existing items for robot and new design items for robots were derived. In addition, 3 ethical design items that should be implemented in autonomous robot were derived.

Online services make users’ communication activities and content public. This online information has contributed to accelerating the creation cycle of user-generated content. Moreover, these services also allow researchers to utilize these online texts as a public source for easily analyzing human activities, also referred to as social sensing studies. However, we need to realize that there exists a controversial problem of privacy especially in the sensitive areas of creation, even though the content is public. This study tries to create new guidelines for online study using the case of the flaming of a study of female fan-fiction, which attempted to extract and filter sexual expressions using online fan-fiction novels as source. Researchers from the fields of both engineering and humanities, including law and ethics, discussed the violations in this case, and extracted ethical, legal, and social issues according to their specific areas of expertise.

In this paper, we consider a k-partition problem with uncertainty of agents' participation. In a traditional k-partition problem, all agent are guaranteed to join the assigned partitions. However, some agents cannot join them in more real-world problem settings. In such a case, some remaining agents have to move from the current partition to another partition to guarantee k-partitions. However, it requires some costs and loses efficiency. Therefore, we need to determine a k partitions by taking into account the uncertainty of agent's participation. In our problem setting, we assume that at most one agent is not going to participate. If one agent cannot participate, we allow one agent among the remaining agents to move from her current partition to another partition. We formalize our problem as an integer programing and show the results of computational experiment.