Although there are various studies on buzzwords, none of them quantitatively define a word's popularity and are insufficient for analyzing trends. This paper defines the trend and fixation of words on Twitter and uses machine learning to predict the duration of the word's trend based on the definition. First, a threshold is set for the number of times a word is used within a certain period of time to define its trendy state. Next, we used multiple machine learning methods to predict whether a word would be in vogue after a certain period of time, based on the changes in the number of times the word was used. We confirmed that it is possible to predict the trend of the word with a high degree of accuracy. Finally, we quantified the importance of each feature of the model and discussed the conditions for a prolonged trend.

Presentation skills are one of the fundamental business skills for people today. However, learning these skills is difficult because people can acquire them only by their experiences. To solve this problem, we present a deep learning based system that can objectively evaluate both oral presentation and slide design and provide feedback for improvement to the users. For the speaking skill assessment, we train a multi-modal neural network including Bi-LSTMs and attention networks to analyze the linguistic and acoustic features of the oral presentation. The proposed network can predict the audiences' 14 types of impressions on the speakers' presentations with an average accuracy of 85.0\%. For the slide design analysis, we have realized a method that can visually analyze the slides independent of the file format and implemented it into the system. It can recognize whether the slide meet the requirement of ten assessment criteria. The average prediction accuracy of the proposed slide model is 81.67%.

This paper presents an investigation into the causal relationships of satisfaction with 5-minute video chats with matched partners on "knew", a recommendation-based online dating service. The "knew" utilizes user registration information and past evaluation to recommend suitable partners. Following a 5-minute video chat between matched pairs, data is gathered on various aspects of satisfaction with the partner, such as overall satisfaction, appearance, impression, personality, and conversation, type of appearance, and so on. Based on this evaluation, the next partner to improve overall satisfaction is recommended. Therefore, comprehending the evaluation mechanism for overall satisfaction is significant. However, this is challenging due to the intricate relationships. The aim of this study is to discover causal relationships between satisfaction with a video chat partner. Utilizing a causal discovery method LiNGAM, we examined the causal relationships between satisfaction types of both genders and found that higher satisfaction with external aspects such as appearance and impression caused higher satisfaction with internal aspects such as personality, conversation, and compatibility, and ultimately higher overall satisfaction for both genders.

Japanese companies are required to advance various corporate reform agendas, including DX and the encouragement of women's success in the workplace. In the process of devising and evaluating their reform initiatives, it is imperative for them to conduct benchmark surveys to discern their standing and identify commendable practices to emulate. However, it is difficult for them to conduct the objective benchmark surveys themselves. In order to provide the companies with a helpful assessment method instead of the manual survey, we developed WISDOM-DX, an automatic assessment system of enterprise activities using a question-answering system based on Web information. To validate the applicability of WISDOM-DX for corporate reform, we evaluated 3380 companies regarding the encouragement of women's success in the workplace through WISDOM-DX. The experiment indicated that 68.1% to 82.3% of the top 10% of companies, as ranked by WISDOM-DX, aligned with three corporate groups that were highly evaluated by experts in regard to women's success in the workplace.

Recently, various high-quality datasets have helped researchers study the microstructure of financial markets as data science. In financial markets, the market order flow is empirically known as persistent: once a buy (sell) order is observed, more buy (sell) orders are likely to appear even in future. This phenomenon is called the long-range correlation (LRC), and its microscopic theoretical model was proposed by Lillo, Mike, and Farmer (LMF) in 2005 based on the order-splitting hypothesis. However, the prediction by the LMF model has not been verified yet because large microscopic financial datasets were unavailable. In this talk (see our latest preprint arXiv:2301.13505), we show that the LMF prediction is empirically verified by data analysis of the Tokyo stock exchange market. Our dataset includes the virtual server identifiers, effectively allowing us to study the microscopic decision-making process of traders. Using this dataset, we tested the LMF prediction and found it holds.

In recent years, an increasing number of companies have been preparing integrated reports as a means of disclosing corporate financial and non-financial information in an integrated manner. Proactive disclosure of information by companies is expected to have the effect of diminishing information asymmetry between companies and their stakeholders. As a result, it is expected that the majority of companies will prepare and publish integrated reports in the future, as they are expected to have an impact on corporate value. In this study, we analyze the current status of integrated reports by market and industry for companies listed on the Tokyo Stock Exchange in Japan. It is assumed that companies in the growth market, which is considered to have high demand for funds, are preparing integrated reports to eliminate information asymmetries and facilitate fundraising.

In current interpersonal communication, information asymmetries arise because the information that each individual possesses cannot be fully shared. In the face of information asymmetry, individuals routinely use the mechanism of "trust" to make decisions and take actions. The relationship between information asymmetry and "trust" is likely to change in the future, as the amount of information in circulation has increased dramatically in recent years due to advances in information and communication technology. In this study, the effect of changes in the relationship between information asymmetry and "trust" on the performance of collective action, an individual and group optimization problem, was analyzed using simulations under specific conditions. The simulation results indicate that collective action performance may be higher when there is information asymmetry or some lack of "trust" than when there is no information asymmetry and complete "trust".

It is important for democracy to realise the right decision about which party to vote for in an election. However, it is not easy to compare the principles of different parties. In response to such situation, the services called "vote matching" have been provided. These services provide candidates of political parties that match to users who answer some questions. In this study, we analysed the data of such vote matching services. From the vote matching data, we analysed the characteristics of voters and their behavioural patterns in response to the results of vote matching.

A scientist's choice of research topic affects the impact of their work and future career. While the disparity between nations in scientific information, funding, and facilities has decreased, scientists on the cutting edge of their fields are not evenly distributed across nations. Here, we quantify relative progress in research topics of a nation from the time-series comparison of reference lists from papers, using 71 million published papers from Scopus. We discover a steady leading–following relationship in research topics between Western nations or Asian city-states and others. Furthermore, we find that a nation's share of information-rich scientists in co-authorship networks correlates highly with that nation's progress in research topics. These results indicate that scientists' relationships continue to dominate scientific evolution in the age of open access to information and explain the failure or success of nations' investments in science.

This study proposes a method for acquiring distributed representations using metapath2vec, which utilizes a network linking social media users and mentions of academic papers. As a result, the acquired academic paper representations that reflect the preferences of social media users and it outperforms citation-based methods in similarity search and clustering, especially for less-cited papers like preprints.

This study proposes a method to measure the bias of users in social media spaces. We model the entire user space of Twitter with a multidimensional Gaussian mixture model using user attributes obtained from Twitter 1\% sample stream data. We then employ the probability of the attribute values range in which users exist, calculated from the model, as an indicator of bias of users in the user space. Numerical experiments show that the bias calculated by the model relevantly reflects the bias observed in the data used to model the user space, suggesting that the proposed method can be used to quantify the bias of users.

In recent years, the widespread use of preprint servers such as arXiv has enabled the rapid release of academic information, which has contributed greatly to the rapid development of deep learning and COVID-19 diagnostic, therapeutic, and drug technologies. However, the quality of preprints is a mix of good and bad because they are not peer-reviewed, so a method to support the discovery of important and reliable preprints is needed. However, it is known that the number of citations used up to now lags behind by years before it increases. In this paper, we analyze the relationship between the number of citations and the number of mentions on Twitter, which is one of the altmetrics, and the time-series changes in the number of mentions of arXiv preprints with high citation counts. Furthermore, we discuss whether the number of mentions can be used for early detection of preprints that will be highly cited in the future.

The demand for professionals skilled in artificial intelligence (AI) is increasing rapidly. As such, it is imperative to assess human resources based on their aptitude, rather than solely relying on self-assessment. To meet this need, we have designed a diagnostic system that estimates applicants' personalities and potential abilities, specifically tailored for the AI domain. Our system was developed using the "POSEIDON" development environment, which facilitates the construction of Bayesian networks and the collection and updating of data. We analyzed assessment items obtained from companies seeking AI talent and compared them with the skill sets published by the Japan Data Scientist Society. Empirical research was conducted at the 3rd Symposium on Industrial applications of Artificial Intelligence (SIAI 2023). To enhance the response rate, we employed innovative designs, such as interactive user interfaces and visualization of results that feature anthropomorphic characters. These methods are not limited to the AI field but can be adapted for use in any domain.

The purpose of this study is to estimate user models for personalized dialogue systems. We consider a method for estimating user models from dialogues using a large-scale language model (GPT-3). Based on the information obtained from the dialogues, a large-scale language model is controlled by prompts to acquire and update user models. User models expressed in sentences are converted into the form of a network. We constructed a dialogue system and verified the effectiveness of the user model acquisition method and the quality of the dialogue.

In recent years, several high-performance conversational systems based on large-scale language models have been proposed, and it is expected to realize long-term conversations by using user information in dialogue history and external knowledge on the Internet. However, previous studies have used dialogue data containing simulated time pauses or first-meeting dialogue data, revealing little about long-term dialogue methods based on changes in speaker relationships and daily events. In this study, we collect long-term text chat data, ranging from the first meeting to eight weeks, and analyze trends and challenges in long-term chats. This paper presents an overview of this text chat corpus and preliminary analysis results.

As the response performance of text-based chat dialogue systems improves, the expectations for the empathy skills of dialogue systems are increasing. Recent research on chat dialog systems focusing on empathy skills has mainly focused on training neural models using a large number of human conversations, and it is not clear what kind of utterances can improve the user's impression of empathy. In this study, we propose a method in which the dialogue system presents its own experiences as a basis for empathy, focusing on "making users grasp the system as an experienceable Other," which is considered important for empathy. As a result of constructing and evaluating several systems with different ways of communicating the basis of empathy, we found that the user's impression of empathy improves when the system uses hearsay as a basis for empathy rather than its own experience as a basis to express empathy

The purpose of this research is to propose a dialogue design for communication robots through non-task-oriented dialogue. The robot's expression of internal states, such as emotions and thoughts, is used for smooth communication and the establishment of close relationships. However, in previous studies in which internal states were expressed by utterances, it has been pointed out that users sometimes do not feel internal states from robots as stated. To solve this problem, we propose the dialogue robot that uses utterances not addressed to another individual (so-called “soliloquies”). The experimental results suggest that soliloquy is a more effective way to increase the degree of anthropomorphism in the robot than talking to users when the robot expresses subjective opinions. Based on these results, we developed an autonomous dialogue android that expresses its internal states including soliloquy and discussed the effectiveness of the system.

Recent models of sentence generation using neural networks have become capable of generating sentences that are as natural as those written by humans. However, the generated sentences may contain factual inconsistencies, which is problematic from a practical standpoint. Studies have been conducted to correct factual inconsistencies using a network that combines a modifier model with another functional model. In such networks, problems have been pointed out, such as errors in one model propagating to the other model. In this study, inspired by GAN (Generative Adversarial Network), we propose a joint learning method that uses corrector to edit factual inaccuracies and a discriminator to detect them. The output of the neural sentence generation model was used to validate the effectiveness of the method.

The purpose of this study is to promote relationship building within a group by a dialogue robot. One approach to relationship building is to increase the information content people know about each other. This study focused on the information content of group members' preferences. We evaluated the effects on relationship building within the group by a dialogue robot referring to common preferences with high information content. The experiment was conducted with the students belonging to the same group interact with the dialogue robot in terms of their willingness to talk with each other and the similarity between them. Based on the results, we discussed the effectiveness of the proposed system.

Discussion capability is important for humans and robots. We have previously developed a discussion system where two robots showed a user discussions about topics with two main claims by using an argumentation structure. However, a method of developing discussion that changes the user's opinion and the effect of robots’ appearances were unclear. In this study, we investigate the effects of discussion development and robots' appearances on user’s understanding. Field experiments were conducted for one month in an exhibition. The results obtained from 2925 conversations suggest that showing interactions where the robot with the opposite stance of the user agreed with another robot with the same stance improved the user's understanding. It is also suggested that when small humanoid robots rather than android robots with the same stance disagreed with another robot with the opposite stance, the user increases the confidence of the opinion.

The research has been investigating humans' expectation of assignment of gender characteristics with robots and its influence on interaction between humans and the robots, and social issues that may be caused by the existence of these gendered robots. The paper shows the overview of the research results, and discusses about conditions necessary for symbiosis of humans and interactive intelligent agents including robots.

When accidents involving AI and robots occur, it is necessary to consider how to resolve disputes. In this report, we clarify comprehensive legal policies to satisfy liability and social acceptability, using the example of utilization in medicine. In conclusion, it is necessary to realize multifaceted measures such as giving people the opportunity to choose AI/robot use, avoiding risks for business operators, and establishing a damage remedy system.

Autonomous Cybernetic Avatars (Cas in short), which act as surrogates for natural persons on the Internet, are beginning to gain importance as an important application area for AI technology in society. Trust between natural persons themselves and CAs, and between CAs and other natural persons or software with which they interact via the Internet, are necessary for their smooth operation. In this report, we discuss this trust from the viewpoint of legal positioning and technical feasibility.

Previous studies have developed theories and estimation techniques based on the premise that emotions are internal states of individuals. However, empathy and the states of spaces and relationships are also important aspects of emotion, that emerge dynamically in the interaction between individuals. We aimed to formulate and measure such "emotions of field." Specifically, the flow of emotions dynamically changing during text-based SNS counseling was analyzed to formulate empathic synchrony. The valence of emotion (negative or not) of each message was estimated at the individual level using a dictionary of emotional words and a mixture-modeling. Regression analysis revealed that the emotional valence of each message was positively predicted by the emotional valence of the message of previous self and the partner, indicating emotional stability and empathic emotional synchrony. Counselors exhibited higher levels of both stability and empathy.

This study aims to acquire a mathematical representation of emotions from sensor data as a data-driven approach to emotion modeling. In order to represent human emotions, a modern representation learning method is used for embedding multimodal expressions into a shared latent representation (an emotional space). The proposed method uses supervised contrastive learning to embed emotionally similar data pairs closer together and dissimilar pairs farther apart, regardless of modality. Human emotions do not fall under any particular category but might be a complex mixture. Therefore, we consider a rating distribution with multiple raters' ratings for a single video, treat it as soft labels, and augment the loss function of supervised contrast learning using the similarity between soft labels. In the experiment using audio-visual data, we evaluate the robustness of emotion recognition when the modality is missing and confirm that the proposed method obtains shared representations of emotions across modalities in a low-dimensional emotional space. We also visualize the emotional space and observe the allocation of non-emotion-related information, such as the gender of the actor, to evaluate the effectiveness of the proposed method in representing the semantic relationships of human emotions.

Motivational Interviewing (MI) is a counseling technique that aims to elicit the Client's (CL) own reasons for behavior change. In MI, positive statements by CL are defined as Change Talk (CT). Previous studies have shown that CL with more CT are more motivated to change their behavior than CL with fewer CT. Other studies have defined the classification of CL utterance labels as a two-class classification problem between CT and Not-CT, and have proposed models for detecting CT using multimodal models of language and facial information. However, there has been no research on CT detection using language, face, and audio information. In this study, we propose a CT detection model that adds audio information to multimodal models using language and facial information. Experimental results show that the addition of audio information does not significantly improve the performance. We also found that weighting by utterance length is effective for Audio information.

Rapport is a harmonious relationship with others. The high rapport between speakers improves the quality of social interaction. We formulate rapport estimation as learning to rank and propose a model that ranks conversational partners based on levels of rapport. This model enables users to re-match with a high rapport partner from a set of people with whom the user has communicated in the past in online language lessons or games using voice chat. Regression models that directly predict rapport ratings can be used to estimate the ranking of conversation partners. However, since rapport annotation is a subjective evaluation, it is biased due to individual differences in perceiver effects, such as response style and positivity effect. On the other hand, the proposed model avoids the problem of perceiver effects by using preference learning, which learns the ordinal relationship between two conversational partners based on the rapport reported by the same user. We compared the proposed model with the regression model using evaluation metrics for ranking. The result indicates that the proposed model is more suitable than the regression model for this task.

We built a humanoid agent system for VR experiences and collected a job interview data corpus. The data corpus includes annotations of interview skill scores graded by third-party experts and self-efficacy annotations by the interviewees, for each question-answer. The data corpus contains various kinds of multimodal data, including audio, biological (i.e., physiological), gaze, and language data. In this study, we developed a feedback system for automated job interview training and analyzed the impact of the feedback. The feedback system utilizes a machine learning model that uses acoustic and linguistic features. In the control group, feedback was provided using a book. The results of the comparison of the effects of the proposed system and the book suggested that the proposed feedback system could suppress the self-confidence of the group that tended to overestimate their performance when compared with the book.

Self-efficacy is a psychological term defined as feeling confident that "I can perform this action in the future. When AI robot systems assist care receivers such as care facilities and hospitals, it is desirable to improve users' self-efficacy by adjusting the difficulty of the target task according to the individual's state. This paper proposes a Kendama task in a VR environment to provide pseudo-successful experiences to model the relationship between difficulty level and self-efficacy. We performed an experiment with 24 participants to investigate the effects of the difficulty level adjustment on self-efficacy. The experimental results suggest that reducing the difficulty level is inappropriate only to improve self-efficacy. Furthermore, it is necessary to increase the difficulty level to leave positive effects on the recall of past and future expectations.

It has been suggested that there is a relationship between arousal level and work performance in various daily life situations. Prior research has identified an inverse U-shaped relationship between physiological arousal and work performance, demonstrating that both excessive decrease and excessive increase in arousal can lead to diminished work efficiency. In this study, we applied the EEG index as the arousal level. A hypothesis was proposed that the capability to naturally regulate arousal at an intermediate level can be enhanced through training with a biofeedback game (BFG). To evaluate the arousal regulation ability, we also developed a visual biofeedback (VBF) system to verify the arousal regulation ability before and after the BFG training. The results indicated that after playing the proposed BFG, the arousal of participants maintained closer to the intermediate level, showing that the proposed game was able to improve the ability to regulate arousal.

Sustaining a sense of adventure is important for improving the likelihood of life opportunities. In this study, we propose a wearable haptic presentation device that promotes emotional arousal with the aim of fostering an emotional state that enhances an individual's sense of adventure. We then investigate the extent to which the user's emotional state is affected by the type of tactile rhythm generated by the device. The results showed that the tactile presentation by the proposed device did not consistently change the emotional valence, but the arousal level changed with the tactile presentation rhythm. Furthermore, the results suggest the existence of rhythmic patterns that enhance cognitive states associated with adventurous attitudes such as "optimistic attitude" and " ease. The present results suggest that tactile presentation with rhythm may be able to change human emotions in a way that encourage adventure.

Embodied AI agents that negotiate with people are needed for proxy negotiation and negotiation training. Various types of information including nonverbal information are exchanged before negotiation so that parties can learn counterpart's preferences and limits, i.e., pre-play communication. However, the relationship between the outcome of negotiation and "emotional pre-play communication", in which parties infer counterpart's preferences and limits from emotional expressions which are considered to be more reliable than verbal information. In the present study, we investigated whether inferring preferences from an AI agent's facial expressions during pre-play communication contributes to an integrative solution in a multi-issue ultimatum game. Participants (n=147) played a multi-issue ultimatum game with an AI agent that expresses its preferences by facial expressions in a (preference learning: present vs. absent) between-participants experiment on-line. The results showed that participants in the preference learning present condition could reach more integrative solutions compared to participants in the preference learning absent condition. This suggests that, in negotiations between humans and AI agents, it is effective to know and model the other party by exchanging nonverbal information before negotiation to reach an integrative solution.

In recent years, robots and virtual AI agents have been equipped with emotion expression functions and have begun to establish social relationships with humans. However, the significance and effects of incorporating emotions into industrial robots, which have a strong tool-like nature, are not fully understood. In this study, we investigated the effects of robot emotion expression on human decision making in a situation where robots assist humans, and how emotion expression interacts with the presentation of information through verbal expression. Participants performed a desert survival task with an arm robot in a laboratory. Participants participated in one of four conditions, which consisted of the presence or absence of emotional expression through the eyes displayed on the LCD and the presence or absence of audio description of the item. In this presentation, we report the results of the experiment.

There are widely used services to look up life science ontologies such as BioPortal, Ontology Lookup Service, and Ontobee. When the author tried them to develop a new ontology and RDF data with it, I found a function that could make them more useful. The function is to search for properties relevant to a given class and vice versa, that is, to search for classes relevant to a given property. Therefore, DBCLS proposes an ontology search service that provides this function. The author has developed Triple Data Profiler, a tool to retrieve relationships among classes and properties used in a given RDF dataset. We assumed the data obtained with this tool can be used to provide the aforementioned service and surveyed them which collected from 41 endpoints and 1003 graphs. Based on this survey, I confirmed the feasibility of this service.

Recently, digital textbook systems and various exercise support systems have been introduced to educational institutions, and research has been conducted to collect and analyze logs in order to improve teaching and learning. In addition, research has been conducted to incorporate group work into exercises in order to prevent differences in the level of understanding in conceptual modeling lectures, and effective pairing methods have been proposed based on the learners' level of understanding. In the previous studies on pairing methods, logs of learning behaviors such as the results of exercises and the status of reading textbooks as well as dialogues were analyzed and surveyed. In this study, we propose a chatbot system that supports learners in preventing the modification of their work and correcting it by incorporating a third party's opinion into the discussion based on the learner's understanding and the discussion situation.

In order to reduce the cost of assembling collaborative robots, we aim to realize a system in which users of collaborative robots can assemble them themselves by referring to an electronic manual on a tablet. To realize this system, it is necessary to provide an interface for the electronic manual that does not interfere with the work. We conducted several experiments in which non-experts assembled a collaborative robot using an electronic manual and remote support from experts. For example, there were errors where non-experts inadvertently skipped reading the electronic manual when they had to repeat similar work. In this research, we summarize these errors and problems with the interface, and describe our plan for improving the interface.

In recent years, using "word embeddings," in which vectors represent word meanings, has made it easier for computers to handle language meanings. However, Word Sense Disambiguation remains an issue for polysemous words. Word Sense Disambiguation determines which sense a polysemous word is used in a sentence. It is an essential task for computers to handle the meaning of language. For Japanese Word Sense Disambiguation, we propose a method to generate word embeddings of words so that the variance between clusters of different word senses is larger and the variance within each cluster is smaller. Our proposed model uses data before and after the target paragraph. The data is paragraphs before and after the target paragraph. We generated word embeddings of five targets word with conventional and our proposed methods, We compare existing and our proposed method for verification. We evaluate the inter-cluster and intra-claster variance and conduct the overall evaluation.

To reduce the cost of introducing collaborative robots, we investigate tablet-based electronic manuals that enable non-experts (robot users) to introduce robots themselves. Since the creation of such manuals is time consuming, to promote the use of robots, a methodology is needed that allows experts (robot manufacturers) to create manuals for non-experts, rather than having knowledge engineers continuously create manuals for each individual robot. In this study, we aimed to establish a methodology for creating manuals for non-experts and conducted a basic study of the elements necessary for a manual, using an introductory manual for a robot as a model. We conducted several evaluation experiments in which non-experts actually assembled a collaborative robot using our prototype manual, and compared and discussed how the success rate of the work changed depending on changes in the manual. For example, we found that the parameters that directly express the success of a task and the tasks involved should be explained in the manual from different perspectives, based on the creation of a manual for teaching procedures for a collaborative robot that must keep errors to a few millimetres.

Tree search is still important in the field of AI for Player versus Player, and AlphaZero combines tree search with machine learning. On the other hand, AI is not only pursues simple performance but also adjusts the difficulty level according to the opponent, is also considered important in servces. What is the most important a fighting style always achieves a desired win rate against the opponent, so AI is needed to achieve a natural objective win rate level Risk-sensitive Satisficing (RS) is algorithm for target-oriented exploration. we proposed AlphaZeRS, which changes the evaluation function of AlphaZero from PUCT to RS. RS feature quick search and discovery to the objective level, which may reduce the number of nodes. In this paper, we tested AlphaZeRS in terms of achieving the target win probability level against opponents of different strengths and saving node deployment through simulations of two-player games.

Selection algorithms for ad delivery and recommender systems have become an indispensable part of Web services. Since the tastes and preferences of people are fluid, to be able to follow them in non-stationary environments is important for those algorithms. We focused on a human decision-making tendency, that is, the tendency to give greater importance to achieving some goal rather than achieving optimization. Agents with this target-oriented tendency are expected to make flexible and highly followable decisions, because they explore according to the degree of achievement without being too sensitive to the changes in the environment. Risk-sensitive Satisficing (RS) is a meta-policy that incorporates target-oriented decision making. Hanayasu et al. showed that RS has excellent followability in non-stationary environments. However, it has not been verified whether it keeps similar followability in non-stationary environments in contextual bandit problems. We used Regional Linear Risk-sensitive Satisficing (RegLinRS), which is an extension of RS to approximate functions, to verify the followability in the environment, and showed the usefulness of RegLinRS.

Humans tend to attempt to achieve higher goals by gradually updating the objective level.Also ,Trial and error to achieve the goal is very quick.These allow for efficient, step-by-step optimization of procedures. The latter trial-and-error capability is supported by the Risk-sensitive Satisficing (RS) algorithm in the context of reinforcement learning. On the other hand, there is a lack of discussion on the step-by-step updating of the objective level in a framework that combines the former with the latter. The advantage of having an objective is that prior knowledge can be used to set the objective. In the case of animals, it corresponds to the search for food using calorie consumption as a minimum criterion, and in industrial applications, it corresponds to operational costs and numerical targets for investors. If the goal is achieved, the agent adjusts the target upward, and if it is unattainable, it adjusts the target downward. It is also very flexible, as it can change its goals based on hearsay information, such as when another agent has achieved a better performance record. In this study, we examine the joint goal search, RS, and the gradual modification of the goal level through simulations of the Bandit problem. We propose a natural form that efficiently optimizes behavior by having an initial objective level corresponding to a prior distribution based on prior knowledge and body structure.

Reinforcement learning is weak to real-world noise and difficult to adapt to the gap between simulation and reality. This problem is famous in motion control tasks and is also remarkably seen in contextual bandit problems used in recommendation systems. Contextual bandit problems require a linear approximation of the target feature, but some algorithms that perform well on artificial data may not be effective for noisy real-world data. Humans adapt dynamically to complex real-world environments with limited data sampling by prioritizing trial and error aimed at reaching a certain aspiration level, rather than optimization. Risk-sensitive Satisficing (RS) is a target-oriented algorithm that includes such human cognitive tendencies.In the contextual bandit problem, RS has been suggested to perform well not only on artificial data but also on real-world data. However, it was necessary to have a certain adoption weighting rate for a prior distribution as a parameter in fitting real-world data. In this study, we tested the possibility of quickly and flexibly adapting to a wider range of data By introducing a meta-algorithm that dynamically determines the adoption weighting rate.

Human beings can achieve a balance between exploration and exploitation by setting an aspiration level, or a goal, and can efficiently learn a behavior sequence which satisfies the goal. Risk-sensitive Satisfying (RS) applies this decision-making tendency to search methods in reinforcement learning. This approach, however, does not work well in all of reinforcement learning settings, since RS does not handle action sequences well. On the other hand, a method is proposed that enables to learn reliability from action sequence. This method draws on experience memory, an approach in reinforcement learning , to compare the current state to the past, and dynamically calculate reliability. This method has followability in unsteady environments, and is expected to surpass the performance of existing methods. On the other hand, its performance has been verified only in a few tasks, still unknown as to its effectiveness in reinforcement learning tasks in general. This time, we verify and discuss this type of future-oriented reliability in various reinforcement learning tasks, and aim at its adaptation to reinforcement learning in general.

For users of hotel booking sites, customer reviews and responses to them from the facilities are important factors for the users’ decision making. Therefore, a system that helps facility staff write back to their customers is highly needed. However, generating adequate responses to negative reviews, which are much fewer than positive ones, is difficult with machine learning approaches, which often assume balanced data. In this study, we attempt to generate responses to negative reviews by controlling responses by using sentiment analysis. We construct a system that combines a sequence-to-sequence model based on GPT-2 and a sentiment classifier based on BERT, and evaluated the system using review data from Rakuten Travel. Through objective evaluation, we show that the system is able to generate more human-like responses. Through subjective evaluation, we show that the model considering sentiment is capable of generating responses that are more appropriate to negative reviews.

Foundation models can be adapted to various tasks by Few-Shot learning, which uses a small number of examples as a prompt. To improve Few-Shot learning, Chain-of-Thought (CoT) prompting has been proposed, which divides the process of thinking into steps. Although the effectiveness of CoT has been proved in English tasks requiring logical reasoning, it has not been verified in Japanese. We examine the effectiveness of CoT in Japanese using a Japanese foundation model, HyperCLOVA JP. We first construct Japanese datasets for the following three tasks: arithmetic, commonsense, and symbolic reasoning. Then, we conduct experiments using HyperCLOVA models of four different sizes. The results showed that CoT prompts were more accurate than standard prompts, and that the performance of CoT prompts was correlated with model size.

Recently, SNSs have facilitated the collection of a wide variety of text data. However, SNS text data has problems such as short sentences with abbreviations and colloquial expressions, which make labeling difficult, and the difficulty of collecting a large amount of data in a short period of time. To solve this problem, data expansion is an effective method for efficiently preparing large-scale, high-quality labeled text data for machine learning. In this research, we aim to improve the learning accuracy of sentiment classification by extending the data to Japanese texts. EDA was used as the data expansion method. In particular, the use of various models for text manipulation in the EDA increased the range of data expansion. The expanded text generated by data expansion was evaluated based on the semantic similarity and the degree of textual change. The optimal data for training was selected by determining a threshold value. The WRIME corpus was used as the dataset to ensure the reliability of the labels. In this presentation, we report the results of learning accuracy in sentiment classification using data expansion.

Causal knowledge is necessary to develop a facilitator agent that can understand the points of discussion and the opinions of the participants. However, causal knowledge contained in Wikidata, a famous knowledge graph, is not sufficient. Therefore, in this study, we have attempted to extend the training data for GPT-3 re-training using Wikidata's causal knowledge as a method for cause extraction. As a result, we had confirmed that the accuracy was improved over the conventional method. In this paper, we hypothesized that multi-stage retraining, rather than mere data expansion, would improve accuracy, and verified this hypothesis through experiments. The results showed that multi-step re-training improved extraction accuracy compared to mere data expansion. Furthermore, we designed a "generality'' scale to determine whether the extracted causes are widely known to the general public, and confirmed the trend that the "generality'' of causal relationships that are widely known to the general public is higher.

Creation of concepts that affect the our life is strongly interesting for organizations such as companies that require innovation. In psychology, generation of an emergent concept from the combination of existing concepts is called conceptual blending, which is often treated as a task of interpreting pairs of nouns in language. On the other hand, since visual information facilitates the interpretation of objects, the generation of visual representations of unknown concepts is an important issue. In this study, we focus on associations, which are considered to be important in creative thinking, and propose a method to realize conceptual blending by incorporating moderate association into the process of combining concepts, and generate images of the concepts from the linguistic interpretation. The effectiveness of the proposed method is partially confirmed by the experiment with human evaluation.

Technology for image recognition has been rapidly evolving recently, and is now applied to many fields including sports science. One long-term goal is the development of systems that analyse player motions when playing different kinds of sports. In this study, we focus on tennis and present a method (RTD3D) that tracks the position of the racket tip in frames of a video captured from a single viewpoint. RTD3D is based on deep convolutional network machine learning. It is trained to take time sequential frames from a tennis video and generate a confidence map for each frame, with a strong confidence peak at the position of the racket tip. To reduce false detections, pre-processing adaptively blurs the background, and post-processing uses a particle filter for tracking, followed by a Hampel filter to improve smoothness. Experiments on a tennis service video show that racket tip detection using RTD3D is more accurate than our previous method. We also show that the pre- and post-processing effectively improves the accuracy and smoothness of the racket tip trajectory estimates.

In recent years, nano-satellites such as CubeSat have been developed by various organizations. Since nano-satellites have limited communication capacity and computational resources, image compression must be performed with reduced computational resources for transmission and reception of captured satellite images. In this study, the designs of computationally resource-efficient convolutional autoencoder that can be installed in a nano-satellite is a compared in terms of recovery accuracy by using different downsampling methods. Three models were evaluated by SSIM and PSNR for images after compression: a model with pooling layers, a model using convolutional layers with wide stride widths instead of pooling layers, and a model using only convolutional layers with wide stride widths. The results showed that using convolutional layers with wide stride widths instead of pooling layers improved the restoration accuracy and better preserved the edges of the image.