This manuscript reports the results and findings obtained by the operation of the dialogue system "V-TA" that is used in a practical training class in Kindai University "Media Informatics Project II". Our 4 month operation revealed the following implications: (i) Cooperation between Students, TAs, and Lecturers accelerates efficient collection of dialogue corpus; (ii) V-TA plays a role as a feedback collection mechanism from students for a class; and (iii) V-TA halves the cost of QA tasks since it can correctly answer questions from students with 56% probability.

This paper describes a sentence style conversion method using recurrent neural network with long short-term memory cells (LSTM-RNN). In the proposed method, LSTM-RNN is used to learn direct style sentences vectorized by one-hot expressions. Then, the sentence end expression of the distal style sentence is removed and the vectorized one is input into the learned model. The next word is predicted until sentence ends, and the obtained word vector sequence is added to the end of the input vector sequence. A direct style sentence is converted by decoding the generated vector sequence into the form of the natural language. We experimented to evaluate the accuracy of the proposed method. As a result, it turned out that a sentence style can be converted by the method.

Automatic text simplification attempts to automatically transform complex sentences into their simpler variants without significantly changing the original meaning. Several researches on automatic text simplification have conducted based on a large-scale monolingual parallel corpus. However, it is costly to manually construct a parallel corpus for text simplification. Therefore, we investigate automatic construction of a large-scale simplified corpus for Japanese from newspaper database corpora. In this paper, we examined several methods for sentence alignment of texts with different complexity levels. Using the best of them, we sentence-align the Mainichi newspaper and Mainichi newspaper for elementary students, thus providing large training materials for automatic text simplification systems.

Recently, fast and massive advertising headline production is highly demanded by growing number of digital ad. Many types of headline generation systems have been developed. But, most of these systems generate headlines systematically by rules and lack generation variety. On the hands, the systems which generate headlines almost randomly satisfy such variety but these headlines are not relevant to ad objective. Until now, it is still difficult to satisfy both variety and relevancy. To this end, we propose Keyword Conditional Variational Autoencoder for advertising headline generation. We regulate generation process by relevant keyword while keeping variety by randomly selected input hidden variables. It can generate variety of headlines and obtain headlines which include a relevant keyword.

Emoji are among the most widely used communication tools worldwide. Because the number of emoji increases every year and there are 82 face emoji, it might be difficult for users to select an appropriate emoji immediately. Moreover, it is troublesome to continue designing new emoji. Therefore, the aim of the present study is to generate an emoji based on input text automatically to facilitate easier communication and eliminate the process of designing new emoji. The proposed model employs conditional variational autoencoders (CVAE), quasi-recurrent neural networks (QRNN) as the text encoder, and the pre-trained word vector GloVe to the embedding layer connected to the text encoder. In the experiments described herein, it will be observed that the proposed method can generate an emoji that corresponds to an input caption, and output image quality is improved using GloVe.

In order to carry out detailed semantic processing for non-grammatical short sentences such as daily conversations, it is necessary to store common sense knowledge. For combining common sense to machine readable dictionary, construction method for manual distributed representation dictionary was proposed based on relationships between words that are derived from traditional ontologies. Moreover, prototype distributed representation dictionary of about 30000 words based on 37 attributes was developed. Manually described distributed representation is readable from both human and machine, and also it has both high scalability and applicability. In the future it is necessary to confirm the objectivity of description by multiple analysts.

We have proposed a novel framework for narrative generation. This framework converts a sequence of states generated by a simulator into the corresponding event sequence in a story world with metaphorical mappings. In this paper, we describe a collection of sets of metaphorical mappings for chess and maze solving. We annotated a part of them with story world calls for the purpose of using them as a training data for scoring drafts of narratives. We conclude that we have 291 sets of metaphorical mappings available for narrative generation.

AR (Augmented Reality) technologies enable us to support shopping in the real world by presenting relevant information about recognized objects. When acquiring related information of objects, it is insufficient to simply search for images or characters related to the objects. It is necessary to select an appropriate information source from multiple dedicated information sources according to the context. Regarding the selection of information source, a selective meta search engine can be used. In this research, we aim to realize a new selective meta search engine considering the context in the real world to take advantage of AR technologies through a development of an AR-based shopping support system. This paper discusses a design of the selective meta search engine for AR applications.

It is difficult to interview real hope from customer through face-to-face consultation. For example, when customers try to purchase houses, even if they explain their hope to the sales staffs, they frequently purchase houses which are different from what they explain before. In other words, getting real hope of customer from face-to-face consultation is a difficult task. As the result, sales staff cannot offer satisfactory houses and lose their customers. In this study, we conducted a questionnaire on customers who bought house to see the relationship between pre-lisning results from them and their actual purchases.

Today, when the point system is sufficiently spread, it is very important to clarify how customers' behavior changes by giving points when each store makes decisions. Therefore, in this research, in order to predict consumer behavior in considering point system, we propose a consumer behavior model considering point system. Then we try to verify the validity of the proposed model by comparing the actual data with the predicted value of the proposed model, and consider the change in consumer behavior due to the difference in point granting method. As a result, it can be shown that the proposed model is valid considering the incentive effect by point granting method based on the discount rate. Furthermore, considering the incentive effect by point granting method based on the discount amount, we are able to confirm the tendency to visit stores that do not cost in time when customers use stores.

Spatial concept transfer learning model was first for the purpose of transferring the knowledge of places acquired in learning environments when the robot moves to new environments. However, in previous studies, this model has not proven to be effective for transferring the knowledge of places to new environments. Therefore, in this paper, we conduct large-scale performance evaluation experiments on name and localization prediction of this model in new environments and we verify whether this model is effective for transferring knowledge of places to a new environment. The experiment results on a larger scale showed that the model has a effectively a high prediction performance of name and location in new environments, and can indeed transfer the knowledge of prior places.

It is generally difficult to recognize and distinguish human motion which are different kinds of motions but whose motion patterns are similar to each other. Conventionally, we focused on the context of motion and proposed a method of alternately performing two processes, motion recognition using context information and re-estimation of context based on recognition results. However, in our previous works, the performance evaluation of detailed algorithms such as reasonable repetition times in these two recognition processes was not discussed. In addition, only an unrealistic ideal distribution is used as the motion appearance probability for predicting motions that may be performed based on the current context, and utility in the probability distribution including noise has not been discussed. In this paper, we aim to investigate these problems and clarify the conditions for performance improvement. Through experiments, a high motion recognition ratio was obtained when the number of iterations was 5 and 10. Furthermore, we confirmed that the proposed method maintains a high motion recognition ratio even if using noisy motion appearance probability. From these results, we have concluded that the proposed method has utility not only for ideal conditions but also for practical motion patterns.

Since sketch is one of the most important creation by human, analysis of sketch drawing process by artificial intelligence is the essential topic. In this paper, we consider a computational model to evaluate the similarity of stroke order. We have proposed CASOOK in order to realize communication between computers and humans. In this research, we are trying to develop CASOOK-SR which is extended CASOOK adopting sketch-rnn. Since evaluating stroke order similarity is a difficult problem, and there was only a method to evaluate the similarity of stroke order by persons. In this paper, we define an objective comparable evaluation function on similarity of stroke order. In our experiment, the correlation coefficient between the function and the similarity evaluation values by a person were 0.344 and 0.446. It is found that our calculation model is the same as human evaluation when an image made from a stroke order matrix is similar to another.

The photo reflective sensor (PRS) is a tiny distant-measurement module whilch is widely used in wearable user-interface. A issue of such wearable PRS devices is the performance degradation when a user data is not included in the training set (we call the inter-user setup). Moreover, the recognition accuracy also degrades when the same user re-wears the device (we call the intra-user setup). In this paper, we introduce the multi-task learning algorithm using the Parallel residual adapters as a method to perform additional training at high speed using a small amount of additional data. In the experiments, we demonstrate the performance improvement in inter-user and intra-user setup with the data of an actual PRS device.

The fluctuation of an RR interval (RRI) on an electrocardiogram (ECG) is called heart rate variability (HRV). Since HRV reflects the activities of the autonomous nervous system, HRV analysis has been used for health monitoring systems. However, the performance of health monitoring systems using HRV features is easily deteriorated by arrhythmias. The present work focuses on premature atrial contraction (PAC) that many healthy people have. To modify RRI data with PAC, the present work proposes a new method based on a denoising autoencoder (DAE), referred to as DAE-based RRI modification (DAE-RM). The proposed method aims to correct the disturbed RRI data by regarding PAC as artifacts. The performance of DAE-RM was evaluated by its application to RRI data which contains artificial PAC (PAC-RRI). The result showed that DAE-RM successfully modified PAC-RRI data. The root means squared error (RMSE) of the modified RRI was improved by 27.4% from the PAC-RRI. The proposed DAE-RM has a potential for realizing precise health monitoring systems which use HRV analysis.

Medical images are widely used in clinical practice for diagnosis and treatment, and much time is spent on diagnosis. Therefore, research to automatically generate cases from medical images has been actively conducted in recent years. However, it is difficult to judge the case as a classification problem because there are orthographic variants in the case written in the medical certificate. In this paper, we aimed to automatically generate character-level cases in order to cope with orthographic variants on chest X-ray images. In addition, the interpretability of the result was improved by introducing an attention mechanism. As a result, it was confirmed cases with features such as position information were generated, and the effectiveness of character-level approach was shown in text generation of medical images.

Patient mix-up on blood samples is one of the common causes of blood test errors. It is also known as patient misidentification problem. Although the detection of mix-up is commonly performed by naive comparison with the last laboratory results of the same patients: delta checks, either the sensitivity or the specificity of delta checks is not satisfactory. To establish a new detection system, we made simulated mix-up data from actual data of blood cell counts (CBC) and serum chemistry in our hospital. Using differences from the previous laboratory results as features, a highly accurate detection system was built by machine learning technique. An XGBoost model recorded the best ROC AUC score of 0.9986.

Esophageal cancer has a 10-year survival rate of about 20%, which is a cancer with a high mortality rate along with pancreatic cancer. It is also known that diagnosis of cancer by CT images is difficult to distinguish between peristaltic movement and cancer stenosis in the gastrointestinal tract such as the esophagus. Therefore, in this study, by performing image recognition learning a CT image of a patient diagnosed as esophageal cancer in the past with a convolution neural network (CNN) and a recurrent neural network (LSTM), it is aimed to construct a system for discriminating the presence or absence of cancer from a new CT image. As a result, we succeeded in a classification model of esophageal cancer using CNN and LSTM, and it to classify with more than 80% accuracy.

We studied whether cerebral hemorrhage can be diagnosed correctly by deep learning with convolutional neural network using CT images taken at our hospital. Data processing techniques such as trimming-off and augmentation improved the precision rate of image classification. Generally, a large amount of image data is required for deep leaning. However, it is often difficult to collect many cases in one hospital. Our results suggest the possibility of on-premise medical image recognition with deep learning in the future.

In the field of image analysis, convolutional neural networks (CNN) have been successful. It is also applied in medical image analysis and includes skin visible images, X-ray images, computed tomography (CT) images, magnetic resonance images (MRI), etc. In this research, we used You Only Look Once version 3 (YOLOv3), one of the latest deep - layer object detection methods, to detect the lesion site and to improve the precision of the previous study for the purpose of using in the medical field. As the verification data, the DeepLesion dataset used in the previous study was used and YOLOv3 was adapted to the chest image in the DeepLesion dataset. A model was created for each CT value used in actual medical practice, and a FROC curve exceeding that of the previous study was obtained. That is, we report reduction of false positive number per sensitivity. In addition, with reference to the fact that the radiologist uses 3D information to identify lesions that are similar to normal tissues, we also report on model modification efforts using neighboring slices of CT images labeled with lesion sites in Deep Lesion.

The purpose of this paper is to estimate people with a high risk of falls and to contribute to fall prevention by clarifying their potential factors. As a result, 6 elderly people with high risk of falling among 98 female participates were extracted using Mahalanobis distance, and 4 out of the 6 elderly people had experienced falls or nearly falls. Furthermore, we extracted 5 factors related to the balance of lower limb muscle strength, the neuromodulation function and so on using factor analysis.

The main purpose of the current article is to survey the general condition of AI technologies utilized in psychiatry. A search on PubMed shows that the number of papers written on application of AI to psychiatric matters has been growing remarkably. “Depression”, “schizophrenia”, and “Alzheimer’s disease” are the most frequent disorders that appeared in those articles published in 2018. “Support vector machine”, “random forest”, and “logistic regression” are the top 3 machine learning models referred there. “Deep learning” still seems to be rather unusual in psychiatric researches. It is definite that AI will bring great progress not only in clinical practices, but also in theoretical bases of psychiatry. Collaboration of AI engineers and clinical psychiatrists and arrangements of the social environment which promotes clinical researches with AI are essential.

Recently, the ECG (Electro Cardio Gram) has attracted to estimate human condition, such as fatigue and stress. Measuring the human biological data for a long time it takes physical and mental load. In this paper, we propose a method to predict long-term ECG data based on short-term data. The typical methods such as LSTM are generally used to predict time series data. The target ECG is characterized by fluctuation in period and voltage, and it is required to predict fluctuating data. So, we evaluated a 1 Dimension Convolution Neural Networks method using a filter size that matches the frequency characteristic of ECG. We showed that the method can be predicted more accurately than the LSTM. This result suggests that it can be an effective means when predicting long-term data.

Investment trust and fund management companies have accumulated a large number of visit records that were summarized by their analysts after conducting hearings against companies. Such visit reports include crucial information of companies such as companies' financial conditions and future strategies, which are used to estimate investment values of individual companies. However, it is not easy even for skilled fund managers to derive suitable market outlooks and investment decisions from a huge amount of accumulated documents. In this research, to support investment decisions, we propose a new LSTM model with self-attention mechanism that can extract important sentences in analyst visit reports. Such extraction is conducted based on the sentence scoring, which is obtained as the weights in a self-attention mechanism. In our experiments for a set of 1,390 visit reports, we demonstrate that the proposed model has about 79% accuracy for extraction on average under the 5-fold cross-validation.

We built the artificial market model (agent-based model) that includes two stocks and one ETF (Exchange Traded Fund) and one arbitrage-agent. We investigated relationship between liquidity and arbitrage costs by the model. Our simulation results showed that when costs are smaller than volatility chances of arbitrage happen more times, number of trades by the arbitrage agent increases and the price difference between ETF and stocks becomes narrower.

In this research, a clustering structure change detection method is employed for analyzing temporal changes of the inter-bank network. The inter-bank network is constructed from data of Italian deposit trading system e-MID, and changes in the structure of the network that focused on the type of the bank are analyzed. By using clustering structure change detection, the change of the inter-bank network's structure after Lehman shock is detected, which depends on the type of bank.

This study aimed to analyze the order behavior by High-frequency traders (HFT) called market making (MM) strategy. We used the order data of Tokyo Stock Exchange provided by Japan Exchange Group, Inc.. Firstly, we preprocessed the order data for merging virtual server used by the same traders. Secondly, we did a cluster analysis of traders based on some indexes indicating features of their trading strategy and extracted HFT-MM traders. Then, we calculated how many ticks their ordering price is far from the last executed price. As a result, we found some of their orders were placed at quite far (5-10 ticks) price from the last executed price for HFT-MM. This result means they mixed some strategies other than market making strategy and the strategies, possibly, will cause the unstabilizing effect when the market price is very fluctuating.

In annual securities report, various information such as results of diverse business performances, point of view about causation of these outcomes, and issues and challenges to be addressed in the near future are included. Most of previous researches proposed the extraction methods of important sentences containing causal information of past company's performances but not effort to address future company's issues from text materials. In this paper, we propose our original method to extract future-oriented sentences by the combination of two SVM identification models, one of which captures features of future and the other aims for purposes and means in sentences of annual reports. All mean evaluations of our models, that were precision, recall and F-score, showed more than almost 0.9. and indicated that by using our model, we can effectively collect future information about business activities from annual reports as well as other relevant sources, which would allow us to make unique investment decisions and to develop unprecedented investment methods.

Predicting expected office rent obtained from a land can be utilized for making investment decisions. Previous research showed that fluctuation of office rent differs in new and renewal contracts. In order to accurately predict office rent, paid rent which is an indicator of rental price in both contracts should be considered as an important determinant of office rent. In this research, posted price is used to calculate the paid rent and estimate the office rent by conducting multiple regression analysis. We present posted price is strongly correlated with rent and effective for prediction of rent.

This paper investigates a hedge and safe haven asset for Bitcoin investors. Bitcoin has been receiving high attention from finance investors because of its high upside return and volatility. The recent finance literature focused upon Bitcoin characteristics as an alternative asset. We take Bitcoin investors’ perspectives and consider how to manage the high volatility of Bitcoin. We employ the definitions of hedge and safe haven based on the finance literature and conduct the respective statistical analyses. Our definition distinguishes a weak and strong hedge (safe haven). Our empirical results show that traditional assets such as global equities and global bonds are weak hedges for Bitcoin. Furthermore, we observe that gold acts as a strong hedge against Bitcoin during an extreme bearish Bitcoin market, although the impact is marginal. There is no strong safe haven asset identified in our data period. Our results imply that the fundamental value of Bitcoin is still unclear, and it is difficult for Bitcoin investors to manage their portfolio risk.

In this research, we conduct an experiment of constructing a causal networks within settlement account briefs. We extracted the causal relations from settlement briefs, and construct a network by connecting them by judging similarities. For calculating the similarities, we use a word2vec model created from the Japanese Wikipedia corpus.We use a method based on a combination of idf values which representing importance of words. In addition, by giving the polarities of causal expression using a polar dictionary, and judgment of synonyms that word2vec can't detect, we define so to speak, "negative relation".

When people are synchronized in their movements - performing the same action at the same time - there can be an increase in positive emotions, for example feelings of camaraderie, similarity, and mutual liking. On the other hand, previous research exploring synchronized movements between humans and computer agents has not found such an effect. In this study, we designed and evaluated a system for exploring human empathy towards computer agents by synchronizing their movements in an audio-haptic rhythm game. Our goal was to explore the mediating effects of two social factors: adjusting the synchronization of the agent’s movements based on the human’s movements and modifying the social presence of the agent. Our results showed that, in contrast to previous research, synchronizing movements by adjusting the agent to the human influenced feelings related to empathy; however, modifying the social presence of the agent did not have an effect.

To investigate human interactive behavior, we designed a drawing interaction task between a human and a recurrent neural network (RNN). This study especially collected a large-scale dataset of drawing for training an RNN and self-report psychiatric questionnaires on the web. We visualized the drawing data and analyzed the questionnaire items collected from 1,035 participants. We found correlations between different psychiatric symptoms and the existence of three-factor solutions on the correlation matrix of all the questionnaires.

Many voice agents are nowadays equipped with spoken dialogue functions. It is assumed that the user’s attachment of a voice agent play an important role in promoting the voice agent. In the present study, we examined whether the degree of politeness of greeting would influence the attachment of the voice agent. Here, we compared effects of the greeting “Sumimasen” (polite “excuse me” in Japanese) with “Ano-“ (inpolite “excuse me”) while participants interacted with a virtual driving assistant system. We found that the polite greeting “Sumimasen” increased the perception of “moderateness” and “frankness.” While, it had little effect on the user’s “likability” toward the voice agent, there were significant correlations between “moderateness” or “frankness” and “likability.” The present results indicate that, by selecting and using greetings appropriately, the perception of politeness may be altered and possibly affect the perception of attachment.

This study aims to improve prediction accuracy by fostering diversity of opinions. We take an approach to give incentive to agents and induce diverse opinions and focus the minority reward system. The previous study assumes that the number of agents is sufficiently large, but the number of agents may be small in real-world situations. We show that the minority reward system is not necessarily efficient if the number of agents is small such as 100. To overcome this drawback, we propose a method to improve the performance by tuning the threshold for determining the minority and show the preliminary result of the evaluation.

Knowledge Graph Completion Challenge 2018, a competition of interpretable AI systems to solve crime story, was held and we participated in it. Our approach is based on discussion agents using IBIS (issue-based information system), a kind of discussion structures. This paper is a part achievement of the design architecture. We design two types of agents: a discussion agent and a facilitation agent. The discussion agents generate hypotheses about the criminate. The facilitator agent ask questions to clarify the detail of the hypotheses. To manage the discussion on the hypotheses, IBIS structure is suitable because it has better interpretability. This approach based on the hypothesis generation has a possibility to be also utilized in the real-world discussion support.

This paper introduces a novel construction strategy in Werewolf Game using reinforcement learning(RL). Were- wolf Game is a type of incomplete information games in which the final results of the game is linked to the success or failure in communication. In this paper, we propose a model that uses RL and estimating other agent’s role in order to learn playing strategy in Werewolf Game. In the proposed model, RL is used for deciding the actions of the learning agent and Naive Bayes classifier is used in order to estimate other agent’s role. Up till now, there is no previous research that has effectively applied RL in Werewolf Game among existing AIwolfs in large scale environ- ments. Therefore, by combining RL and estimation of other agent’s role, we demonstrate through experimentation that the proposed approach achieved high level of performance in 11 people Werewolf Game.

Intention recognition is the task of inferring the intentions and goals of an agent. Intention recognition has many applications. Especially, it can be very useful in the context of intelligent personal assistants like robots or mobile applications, smart environments, and monitoring user needs. We present a method to infer the possible goals of an agent by observing him in a series of successful attempts to reach them. We model this problem as a case of concept learning and propose an algorithm to produce concise hypotheses. However, this first proposal does not take into account the sequential nature of our observations and we discuss how we can infer better hypotheses when we can make some assumption about the behavior of the agents and use background knowledge about the dynamics of the environment. Then we talk about future work to improve our method.

We have developed a new decision-making system that combines behavior trees and state machines into a single system. The system has both flexibility of behavior tree and strict control of state machines to give a scalability to development of a character AI. The new decision-making system, we call the AI Graph, extends the node formalism to enable sharing nodes between FSMs and Behavior Trees, provides advanced techniques for code reuse using trays which organize code reuse and behavior blackboards, and also provides many features for integrating with detailed low-level character behavior

For dealing with continuous state spaces, DQN and other algorithms have been proposed in reinforcement learning (RL). However, it is hard for DQN to explore efficiently, as it depends on random search strategies such as epsilon-greedy. Humans are known to effectively search and learn through "satisficing" instead of optimizing. Although the risk-sensitive satisificing (RS) algorithm enables satisficing in RL, it depends on the count of visiting each state, which poses a problem for continuous spaces. We propose a method for solving this problem by pseudocount and hash+auto encoder methods that enables intrinsically motivated exploration. Through two experiments, we show that RS combined with the two methods enables deep satisficing RL that searches and learns efficiently in continuous spaces.

We consider a prediction-based decision-making problem, in which a binary decision corresponds to whether or not a numerical variable is predicted to exceed a given threshold. The final goal is to predict a binary label, however, we can exploit the numerical variable in the training phase as side-information. In addition, we focus on class-imbalanced situation. We investigate on an idea of using near-miss samples, which is specified by the numerical variable, to deal with the class-imbalance. We present the benefit of exploiting the side-information theoretically as well as experimentally.