Since the Internet has become popular in the world, electrical commerce(e-commerce) has made a major advance and established an enormous market. According to Ministry of Internal Affairs and Communications in Japan, it is reported that global sales of e-commerce reached 4.25 trillion USD in 2020 and expected that the global sales keep increasing. In order to acquire customers, e-commerce companies have been taking various actions. Discount price coupon is one of the effective ways for customer acquisition. Previous studies show that coupons increase purchase intention. However, they have not only positive side. Compensation for discount price reduces revenue. In addition, coupon targeting is limited by budget. For these reasons, we set our goal to invent a cost-effective coupon targeting method under budget constraint. In this article, we proposed a cost-effective coupon targeting method based on uplift modeling. The proposed method decides coupon users with an indicator of cost-effectiveness estimated by machine learning. By coupon targeting simulation and comparison with a conventional method, we verified that the proposed method outperforms the conventional method.

Large language models (LLM) have an outstanding ability to accurately generate documents and code following the intent of brief instructions input by humans. In the future society, it is expected that ideas conceived by humans will be immediately materialized through LLMs, thereby increasing the importance of continuous and rapid generation of ideas by humans. However, continuously producing product improvement ideas from various perspectives, such as business strategy and usability, is a challenging task for humans. In this study, we aim to develop a technique that automatically recommends improvement ideas for existing products, mainly software products, based on data resources. The proposed technique repeatedly analyzes data and creates improvement ideas through LLMs with prompt templates defined based on the requirements process model. Although the experimental results with ChatGPT did not demonstrate the effectiveness of the proposed technique, they highlighted several new challenges for future study.

In the marketing field, online service providers have recently implemented personalized recommendations on various mobile applications as a part of customer relationship management. However, there has long been an issue of consumer heterogeneity, where each customer has internal differences that are difficult to discern from behavioral logs. On the other hand, the transfer of pre-trained model referred as large-scale language models (LLMs) has facilitated text data analysis, such as customers' reviews on the online platform, wherein they express their reasons for the evaluations which cannot be obtained from behavior logs. Therefore, in this study, we introduce a conceptual model of multimodal deep learning, combining review texts with traditional customer and store information. To rephrase, we develop a restaurant evaluation model that integrates text data analysis to comprehend consumer heterogeneity, alongside conventional analytical methods. Our comprehensive exploration and comparison of multiple models reveal that the proposed model shows the best prediction accuracy. Moreover, we discuss the limitations of analyses relying solely on traditional customer information and the potential for advancements in future purchase prediction models.

In recent years, the significance of addressing food loss has increased due to its environmental impact, economic loss, and concerns about the sustainability of food supply. This is especially relevant in retail stores, where large amounts of food loss occur, and its reduction is highly sought after. In this study, we developed a discrete-time display quantity optimization algorithm based on demand forecasting to tackle these challenges. This algorithm learns demand trends from past sales data and suggests the optimal display quantity to minimize food loss from excessive procurement and loss of sales opportunities. The optimal display quantity is output as a single value per unit time, eliminating the need to track detailed display status during operation. This makes the algorithm highly practical. In addition, we formulated optimization requirements and constraints in store operations and validated the effectiveness of the algorithm through numerical experiments and trial applications in actual store operations.

Obtaining users' preferences during a dialogue is desirable to provide personalized services. We collected interview dialogues aimed at acquiring food preferences, and created a response generation model based on the intention and semantic content of the interviewer's utterance by fine-tuning GPT-3.In this study, we investigated the performance of the proposed model by comparing with ground truth interviewer utterance, Zero-shot ChatGPT and a fine-tuned GPT-3 model that directly generates only response sentences as baselines. The subjective evaluation showed that in terms of eliciting the interviewees' food preference, the proposed model's response sentences were superior to those of the baseline models and comparable to real human interviews.Analysis of the characteristics of the response revealed that the proposed method 1) frequently generates questions in various dialogues and 2) produces more detailed and context-related questions compared to ChatGPT.

It is known that Japanese onomatopoeia has an aspect that has a semantic contrast between voiced and voiceless consonants (e.g., giragira-kirakira) at the beginning of the word. In this paper, we analyzed the difference between voiced and voiceless consonants of onomatopoeia in a space of static word embeddings such as Word2Vec. In an experiment in which we classified onomatopoeia starting with voiced or voiceless consonants in the space of word embeddings, we got a classification accuracy of 0.84 at best, which showed the possibility that word embeddings have some information about the sound symbolism of voiced and voiceless consonants. In an experiment in which we compared the contrast of voiced-voiceless consonants and bipolar adjective pairs, we got a result that showed that particular adjective pairs (e.g., "beautiful-ugly", "bright-dark") have some correlations with the contrast of the consonants, and the correlations have some compatibility with results shown by previous research.

This paper discusses experiments designed to answer the question, "How do neural language models learn iconicity in natural language?" We implemented a regression task to predict iconicity ratings of English words using a model based on English BERT. Despite the model not being explicitly informed about the word classes (e.g., noun, verb, onomatopoeia), it demonstrated a tendency for the predicted values to vary according to word class. Identifying the factors contributing to these findings and extending the experimentation to models in other languages represent essential future work.

Large language models (LLMs) are flexible and can handle various natural language processing tasks. Many spoken dialogue systems are realized by linking a dialogue model built using an LLM with other modules, such as speech recognition or synthesis systems. However, such a cascaded model with multiple modules is complicated and tends to propagate errors from the previous module. The model can also not consider sensitive expressions in the non-verbal representation of dialogue because the discrete representation, such as texts, is used to connect modules. This research aims to solve these problems by converting the input speech into a vector of continuous expressions and connecting it to a dialogue model. The experimental results show that the generated sentences do not fully take the dialogue context into account, and there is room for improvement, but the natural sentence generation is learned, suggesting that a dialogue model using continuous expressions is feasible.

In recent years, research on dialogue agents has been actively conducted. One of the previous studies is a method that decouples strategy and generation using dialogue act that represents the intention of the utterance. This method using dialog acts in the sentence generation process improved the task success rate and the human-likeness of the utterances. On the other hand, since the parser is implemented rule-based, there is a limitation to the sentences it can parse the dialogue acts. Based on the above background, in this study we annotate training data based on the proposed dialogue acts and propose a parser based on deep learning. The parser using deep learning shows that the accuracy of classification for dialogue act is approximately 83%. Furthermore, we succeeded in reducing the number of unknown dialogue act.

To use sentence generation AI safely and securely, it is necessary to show evidence citing the literatures. Therefore, in this work, we propose a system to show evidence for reason of dangerous behaviors in home generated by sentence generation AI for the dataset presented by Knowledge Graph Reasoning Challenge 2023. First, we extract dangerous behaviors in home. Second, Sentence Generation AI generate reasons of it. Third, Retrieval Augmented Generation (RAG) retrieves a sentence similar to this reason from the literatures on dangerous behaviors in Home and shows the user as evidence. We did a survey to evaluate whether the sentence generation AI can appropriately generate reasons for dangerous behaviors in the home, and whether the evidence showed by the proposed system for these reasons is appropriate. As a result, average score of five-stage evaluation were 3.6 and 2.6. It is found that the proposed system can show general evidence for reasons.

This study develops a system to visualize accessible regions for childrens where accidents can occur in indoor environments using a posture estimation function of digital human models considering both kinematics and mechanics. For parents raising children, visualizing the potential hazards inside their own home can help them improve the environment to prevent inadvertent accidents. Our system simulates three-dimensional interaction of the children with the entities in the home environment using a digital human model that statistically reflects the physical characteristics of children. Reaching posture of the child digital human model that represents each age was estimated by solving inverse kinematics considering mechanical stability. Potential risk areas due to the reach of children were visualized by coloring the environmental model.

Caregivers need to understand the daily living activities of elderly people to design living environments that meet their diverse needs. This paper proposes a system that automatically summarizes the daily activities of elderly people without defining in advance the behaviors to be analyzed, based on a motion-based analysis focusing on interactions between the elderly people and objects in the living rooms. To evaluate the effectiveness of the developed system, it was applied to three months of behavioral data of five elderly people in their 80s residing in nursing homes. The experimental results demonstrate that the system can detect abnormal behaviors such as repeated unnatural drawer use. In addition, this report presents that the system can detect behavioral changes due to installing new furniture in the environment by comparing daily activities.

We held the final presentation of the 2nd International Knowledge Graph Inference Challenge on February 8, 2024 as a workshop in conjunction with the International Conference on Semantic Computing. The main task of the Challenge was to obtain statistics about actions, objects, and locations from videos and knowledge graphs generated using a 3D simulator of parts of daily life. The unique feature of this challenge is that it provides data with a missing part of the knowledge graph, and it is necessary to compensate for the missing information by extracting information from the video and predicting using machine learning on the knowledge graph. In this presentation, we provide an overview of the dataset and tasks of this inference challenge and introduce the four submissions. Since several of the submissions used multimodal LLMs, we will compare them and also discuss the challenges and expectations for current multimodal LLMs in this task.

Text commentary on sports websites is manually entered by sports data companies. It’s very convenient for those who cannot watch the live game but want to follow it in real time. However, making it requires a lot of work, especially in soccer, where a large number of events take place in a 90-minute period by players. As a labor-saving method, we propose an AI-based automatic text commentary generation system using Play data which has a time series of events made of numerical data. We compared the capability of two methods, an NN based method proposed in a previous study and a newly proposed Rule-based method, by Automatic evaluation and Human evaluation, respectively. The results show that the Rule based method significantly outperforms the NN-based method in terms of content selection, grammatical accuracy, and proper noun output, indicating the effectiveness for Data2Text task.

The Knowledge Graph Reasoning Challenge for Social Issues (KGRC4SI) 2022 was held in Japan. The challenge aimed to encourage the development of systems capable of identifying and explaining dangerous situations that might occur in the homes of older people. To achieve this goal, the organizers provided videos simulating daily activities with the household simulator VirtualHome and knowledge graphs converted from the videos using VirtualHome2KG. The primary task of the KGRC4SI was to identify dangerous situations from the provided video and knowledge graph. However, much data representing daily life situations is needed since not all videos and knowledge graphs include dangerous situations. Creating action scripts manually is not intuitive, and it will be easier if users can communicate their intentions and requirements using natural language statements. We have proposed a system for generating action scripts from descriptions of daily life activities using GPT-3.5 Turbo, one of the LLMs, and evaluated their similarity to correct data and execution rate. This research evaluates the system using GPT-3.5 Turbo, GPT-4, and Llama 2 and compares and discusses the results.

An interactive navigation system for VirtualHome has been proposed to guide users to places and things they need by inferring their potential requirements from their utterances based on commonsense and home environment knowledge graphs. In the previous work, the cost of constructing knowledge graphs and dialogue rules was high. We propose a system that has the same functionality as previous work using GPT (GPT-3.5 and GPT-4). We designed prompts for navigation and inference of users' potential requirements by extracting the relationships between rooms and objects in the VirtualHome environment. We also designed prompts to determine the timing of guidance based on utterances. By providing these prompts to GPT, we realized the main modules in the proposed system. Using the evaluation datasets, we compared and evaluated the proposed system with the system of previous work. We found that the proposed system outperformed the accuracy of the system of previous work.

Entity linking (EL), a task associating named entities in text with entities in a knowledge base, has attracted attention as a fundamental technology for question answering and other applications. Most existing EL methods focus on English and may not support other languages or have poor performance. In this study, we propose an EL method for Japanese and English based on GPT, which has advanced language understanding and generalization capabilities. Our approach extracts entity names and generate corresponding Wikipedia URLs from EL target sentences by providing prompts to GPT-3.5 Turbo and GPT-4. Subsequently, we query Wikidata's SPARQL endpoint to obtain Wikidata IDs from Wikipedia URLs and outputs the sets of entity names and their Wikidata IDs. We compared our proposed method with a prior research method (PNEL) on LC-QuAD2.0, SimpleQuestions, and WebQSP datasets in Japanese and English. Results showed that our method outperformed PNEL on all datasets except Japanese SimpleQuestions.

In this paper, we explore a two-period matching problem in which agents' preferences may evolve over time. This model enables agents to be matched with different partners over the two periods. Previous research defines dynamic stability as the absence of any single agent or pair of agents gaining an advantage through coalition across any period. We introduce new fairness, nonwastefulness and individually rationality notions tailored for the two-period matching setting and examine its relationship with dynamic stability. We also discuss the existence of matching that simultaneously satisfies the newly introduced fairness and nonwastefulness properties.

A new matching mechanism under relative ditributional constraints. In this paper, we consider a many-to-one matching problem under relative distributional constraints. We propose a new many-to-one matching mechanism that flexibly determines a matching based on the sequential dictatorship mechanism. We theoretically show that our mechanism satisfies strategy-proofness. Furthermore, experimental results show that our proposed mechanism can improve efficiency compared with an existing matching mechanism with relative distributed constraints.

A combinatorial auction is a type of auction where bidders can place bids or assign values to different subsets of goods (bundles). One of the challenges in combinatorial auctions is the exponential increase in the number of possible subsets of goods as the number of items grows. Traditionally, it is assumed that bidders bid on every possible subset of goods, but this becomes impractical with a large number of items. To address this issue, recent developments include the design of iterative combinatorial auctions using machine learning (ML). This method involves estimating each bidder's valuation function with ML models by asking them to value bundles, thereby seeking a more efficient allocation progressively. However, existing research typically involves training separate models for each bidder, which can be inefficient in cases where there are many bidders with similar valuation functions and a limited number of queries. In our study, we have attempted to design a more efficient iterative combinatorial auction using multi-task learning, which allows for the sharing of information between models. This approach has proven to be more efficient in scenarios with many bidders or multiple bidders with similar valuation functions compared to existing studies.

It is known that there is no stable algorithm that satisfy strategy-proofness in the two-sided matching with information diffusion. In this research, we examine the existence of stable algorithms by weakening the incentive requirement. More precisely, we show that even the non-obvious manipulability property is not achievable by stable algorithms.

This paper organizes transaction data from the floral market and discusses estimation methods for visualizing the demand for chrysanthemums. Due to the extended time from cultivation to harvest for flowers, managing supply appropriately becomes challenging, leading to price instability. Therefore, forecasting demand is essential to regulate supply and stabilize prices effectively. The paper addresses the organization of transaction data in the floral market and utilizes multiple regression analysis to reveal the relationship between price and transaction data.

With the rise of text/image generation models, there are many efforts to utilize them in digital game production. However, most of these efforts are limited to a small part of the text/image data production process performed by humans. In this research, to verify the possibility of fully automatic creation of game content using generative models, we build a system called Red Ram that creates mystery adventure games on demand. Red Ram automatically creates almost all game content, including the story, character images, and conversation scenarios, based on simple instructions from the user regarding the story. We explain the flow of Red Ram's game content creation using text/image generation models and the system construction procedures. We also discuss the current issues facing Red Ram that we discovered as a result of generating game content.

Automatic generation by AI has the great advantage of obtaining a large amount of similar output at low cost. In addition, works with looping structures, which have become popular in recent years, are characterized by the repetition of similar developments. In this study, we attempted to develop a looping 2D adventure game system by combining several techniques of automatic AI generation in order to develop a game system that makes the best use of AI that can generate a large number of similar outputs. Specifically, we developed a system that integrates automatic generation of map design, scenarios, sound, and character animation with reinforcement learning to change the behavior patterns of enemies. By combining these elements, we constructed a system for a looping 2D adventure game. As a result, for example, in the automatic generation of scenarios, efficiency was achieved by producing a large number of plots for each loop by generating outputs that have similar developments but differ in details. We have shown that it is possible to produce games by integrating the results of generation using the above multiple AI automatic generation technologies.

The AI (character AI) used to control non-player characters (NPCs) in entertainment games is required to behave as determined by the developer, rather than pursuing the best actions according to the game rules. Therefore, the current mainstream of character AI is to use symbolic AI methods such as behavior trees to make it easier for developers to understand and control behavior. However, when introducing reinforcement learning to game developments, there may be issues such as the difficulty of alignment and changes in the game environment during the development. In this study, to solve this problem, we propose a new behavior control method that combines symbolic AI and LLM. First, the NPC's specifications, in-game situation, and behavior are expressed in text, and behavior evaluation data is generated by evaluating these using LLM. The generated evaluation data is used as annotation data to build small-scale machine learning models that can run in real time. In the experiment, we prepared behavior rules and policies for the character AI in an actual game, created a model using this method, and validated that the AI can select actions based on the requirements set in the game.

Persona-aware dialogue systems can improve the consistency of system responses, user trust, and enjoyment. Filtering nonpersona-like utterances is important for constructing these systems. We proposed a novel model to capture the intensity of persona characteristics in a given utterance. We trained the model using contrastive learning, based on the similarity of the utterances' speaker. Experimental results showed that our model outperforms existing methods and baselines using pre-trained language models.

To explore the social applications of reinforcement learning, research and development of AI capable of playing games that simulate the complexity of the real world is beneficial. However, opportunities to learn from such complex games are rare. We have developed an AI learning interface for a widely played video game that possesses a considerable complexity. To demonstrate the feasibility of reinforcement learning through this interface, we have developed AI which can play the game with reinforcement learning, and the result indicate that the AI can handle the game's complexity. Furthermore, this effort showed the potential to bridge AI and the general public.

In recent years, with respect to various digital games, there is a method of artificial intelligence that is similar to human ability and more. this method is deep learning, which performs highly in image recognition, time-sequence analysis, or natural language processes. Especially, in the game field, the deep reinforcement learning method is often applied and a gameplay result is more than human performance. However, there is a possibility that humans will act in ways that humans deem inappropriate or have difficulty interpreting the thinking process. On the other hand, a Language model is applied to the digital game field and gets high-performance results. we suppose that this model makes it possible to interpret the thinking process. In this study, the construction of a machine learning model to apply to the agent model in the simulation space is the goal. In this report, the preliminary evaluation goal is to propose the agent model to apply the machine learning method with a time sequence analysis process like the language model in the simulation space.

Many municipalities and health insurance associations hold walking events to promote health, and many of them offer incentives. In recent years, NFT has been attracting attention, but the current situation is that sales profit from secondary distribution is the main objective. If the effectiveness of NFT as an incentive that does not aim at sales profit is clarified, it can be expected to be used in walking events by public organizations. Therefore, in this study, the acceptability of NFT as an incentive that does not aim at sales profit was investigated by conducting a stamp rally. As a result, the data showed that NFT is accepted as an incentive regardless of generation, even if it does not aim at sales profit through secondary distribution. This paper introduces a case study in Kurume City.

Several attempts have been made to implement text command control for game agents. However, current technologies are limited to processing predefined format commands. This paper proposes a pioneering text command control system for a game agent that can understand natural language commands expressed in free-form. The proposed system uses a large language model (LLM) for code generation to interpret and transform natural language commands into behavior branch, a proposed knowledge expression based on behavior trees, which facilitates execution by the game agent. This study conducted empirical validation within a game environment that simulates a Pokémon game and involved multiple participants. The results confirmed the system's ability to understand and carry out natural language commands, representing a noteworthy in the realm of real-time language interactive game agents.

Serendipity-oriented recommender systems have been proposed to prevent over-specialization in user preferences. However, evaluating serendipity is challenging due to its reliance on subjective user emotions. We try to address this issue by leveraging the rich knowledge of large language models (LLMs), which can perform a variety of tasks. As a first step, this study investigates the alignment between serendipity assessments made by LLMs and those by humans. Specifically, using GPT-3.5 we assess the serendipity of recommended items based on users' evaluation history. We evaluate the accuracy of assessment made by the LLM using an annotated benchmark dataset. Experimental results indicate that our method outperforms the baseline method, showing improvements of up to 0.6, 4.9 and 1.5 points in Accuracy, Precision and Macro-F1-score.

This paper proposes a method for constructing personal value-based user models from review texts using LLM (Large Language Models). The RMrate (Rating Matching Rate) has been proposed as a metric to quantitatively assess the intensity of user preferences towards item attributes when selecting items, and has been applied to personal value-based models. Although its effectiveness in information recommendation has been demonstrated, existing methods require explicit attribute evaluations. To address this issue, the proposed method calculates RMrate by applying LLM to extract the evaluation polarity of item attributes mentioned in reviews through prompting. This paper conducts experiments with movies as the target items, demonstrating the accuracy of extracting evaluation polarities and the effectiveness of the proposed method when applied to a recommendation system.

In this presentation, we highlight the methodological aspects of our latest work published in Tourism Management (https://doi.org/10.1016/j.tourman.2023.104821). In this work, we used over 1 million Reddit postings from January 2018 to January 2021, selected 3093 authors in three periods: 1. Before Covid 1, 2. Before Covid-2, 3. During Covid, and classified the authors based on psychological attributes for their posts in the first period. We created word vectors describing two psychographic characteristics: “openness to change" and "conservative" based on the Basic Human Values theory. By use of the word embedding technique, we classified these authors into the two groups by calculating semantic similarities between their postings during the first period and the two respective word vectors. Our results showed that open-minded authors had more positive travel sentiment in the third period than conservative authors, while conservative authors increased risk awareness in the third period compared to open authors. Our study emphasizes that by classifying the authors of large-scale data based on psychological attributes, it is possible to predict the attitudes and behaviors that authors will express in the future, and that the application of theories in psychology and social sciences can deepen the insights obtained from large-scale data.

Information about tourist destinations is usually disseminated by the residents of the area. However, there often exists a gap between this information and the attractions felt by visitors. Understanding the factors that cause this gap and finding solutions are important. How to narrow or eliminate this gap in the dissemination of information about tourist destinations, and specific methods of improvement are still not clear. The purpose of this study is to identify the differences in attractions between official tourism information and visitor reviews, and to propose a method to clarify these differences. For the relevant data, similarity is calculated using a Transformer model. The differences from the group of official information that are similar are explained by the Transformer model. Specifically, the vectors of the documents are normalized, and similarity is calculated using the dot product. The top 10 documents with high similarity are selected, and sentences that generate the commonalities and differences with specific reviews are created. These sentences are evaluated by experts in the tourism industry. A tool for improving the dissemination of tourism information was developed and its accuracy was evaluated. This made it possible to identify areas for improvement in information dissemination at the target tourist destination.

In recent years, the demand for data analysis has increased, and with it, the environment for analysis is required. The basis of data analysis is comparison, and many studies have focused on the comparison of multiple text data in text analysis. In this case, a method is used in which characteristic words in the text are extracted and used for comparison. However, with this method, the perspective of the extracted words may not match, and it is difficult to use features with different perspectives for comparison. Therefore, the purpose of this study is to construct a system that outputs differences and common points simultaneously with viewpoints using ChatGPT's API, and to enable comparisons with matched viewpoints. As an experiment to evaluate the system, several prepared pairs were input into the system, and the validity of the results was evaluated from multiple perspectives based on questionnaires to the subjects.

The PBL exercise is a class for students to learn the essentials to become a data scientist. In collaboration with a company, students work in groups to discover issues and make proposals using real data. Through PBL exercises, students are expected to acquire the ability to analyze data and communicate within a group. On the other hand, various theories of teaching and learning have been proposed over the years in the field of educational technology, and attempts have been made to construct an ontology of these theories. In this study, using the PBL Exercise I conducted at the Faculty of Social and Information Sciences of the University of Hyogo as a motif, we clarify each learning step that constitutes the PBL Exercise I class and examine which theory of learning and teaching applies to it. Furthermore, each teaching theory assumes KPIs, and we will confirm that each learning step actually improves each KPI. The aim of this study is to provide theoretical support for PBL Exercise I, which tends to emphasize practical aspects and effectiveness.

Japanese universities are required to foster data scientists with mathematical thinking and data analysis skills. However, there is no established teaching method to foster data scientists, and it is difficult for university teachers, who are not required to have a teaching license, to evaluate and improve their classes. In this study, we propose a methodology for modeling teaching by using functional decomposition trees used in ontology engineering and combining teaching and learning theories that have been studied in the field of education in the past. Based on the proposed methodology, we also discuss the advantages and problems of the proposed method by visualizing the relationship between theory and practice through actual modeling of PBL-style exercise-type classes conducted at the University of Hyogo.

Analyzing surveys utilizing open-ended responses to questionnaires is a valuable approach to elucidating respondents' perspectives and opinions, thereby gaining insights. However, the analysis of responses on a large scale necessitates a considerable amount of manual labor. Thus, this paper takes an approach of automating the analysis of open-ended responses using large language models. We have generated several types of pseudo data for training category classification models and evaluated the performance of the models trained on each dataset. Through this process, we examine the performance improvements of category classification models using the pseudo datasets automatically generated and annotated by large language models. Evaluation results show that, through several operations of pseudo open-ended responses, we improved the category classification performance against real open-ended responses from 77% to 83%.

In recent years, YouTube has become widely used as a platform where people around the world share information and gather a variety of content and a wide range of user opinions. However, the volume of comments on videos is enormous, and manually analyzing important opinions and trends is inefficient and time-consuming. Therefore, this study proposes a method to analyze YouTube comments using ChatGPT and automatically extract key opinions and emotional tendencies. This will enable viewers to quickly and efficiently understand reactions and opinions about the video, helping them to decide whether to watch it or not.

In recent years, the opportunities for data analysis have increased, and the integration of diverse information requires an enormous amount of time and effort. In addition, the emergence of large language models has made it possible to automate sophisticated tasks that were once thought to be difficult to automate. However, in automatic opinion aggregation, the traditional method of opinion extraction is the mainstream, and automatic abstraction of opinions is rarely seen. In this study, we propose a system that automatically aggregates multiple comments by abstracting them using ChatGPT. We aim to create an environment in which users can easily acquire generalized knowledge from multiple comments. Through preliminary experiments, we compared the automatically generated summaries by the system with those by humans, and evaluated the quality of the summaries.

Generative AI has been attracting attention in recent years, and chatbots and smart speakers are two of the ways in which it is being used, and the opportunities for their use are increasing. Chatbots have pre-defined personalities that cannot be changed by the user. By providing a person's lines and other relevant information as input, the output can reflect the way the person speaks and thinks, allowing for user-preferred dialogue. Many studies have focused on the way a person speaks using dialogue information. However, they do not focus on the generation of person images, and the content of the learned person images cannot be confirmed. In this study, we propose a system that extracts lines that reflect a person's characteristics using ChatGPT and adds an explanation of the situation in which the lines are used, thereby generating a portrait of a person from a smaller number of lines, consisting of the person's actions and the ''way of thinking about things,'' ''policy,'' and ''action policy''' that form the basis of the lines. Through experiments, we have changed the method for extracting lines. Through experiments, we compared the estimation results with different methods of extracting lines, and evaluated the quality of the estimation results.

In recent years, the number of users of social networking services has been increasing, and with it, opportunities for communication with others. While this is a good thing because it makes it easy to communicate with others, messages and posts that hurt or offend others are on the increase. The offensive contents include not only objectively understandable contents such as slander and libel, but also contents that the other party may find offensive depending on the relationship between the two parties and the situation. There are few existing communication support tools that directly correct input comments. In this study, in addition to objective criteria, we use ChatGPT to extract features of sentences that the other party finds favorable or unfavorable from the dialogue history between the user and the other party. We aim to realize better communication by constructing a system that can propose sentence modification based on these features to the other party, if necessary.

This study aims to realize the information access method based on story concepts. A story concept is a short description of the bones of a story structure. For example, "This story is about a man and a woman who belong to each other's unfriendly groups fall in love." In our previous study, we collected short summaries of respondents' favorite stories. Analyzing those summaries revealed that they can be used as story concepts. In this paper, we used the collected short summaries as story concepts, comparing them with outlines of stories collected from the Internet, to clarify features of story concepts: We made sentence vectors from story concepts and outlined each, then examined differences between the story concepts and the outline of the same work by calculating the similarity of their vectors. The result of the study suggested that a story concept can be made from the outline.

Zero-shot classification will produce different classification results for the same texts depending on the input label set.In this paper, we propose a method to generate a large number of candidate label sets for the same zero-shot classification target by antonym substitution and conversion to synonyms using WordNet and find appropriate labels from themFour zero-shot classification methods are evaluated: 1. cosine similarity of text by BERT model, 2. cosine similarity of text by OpenAI's model, and 3. pre-trained zero-shot model of MoritzLauer.In the evaluation experiment, we collected 50 listening test dialogues from each of the N5 to N1 levels of the past Japanese Language Proficiency Test (JLPT) and classified them manually.Three classification attributes of Dialogue Location (6 categories), Speaker's Relationship (2 categories and 4 categories), and Dialogue Style (2 categories) were evaluated.We prepared 212 candidate label sets and counted the RMSE (Root Mean Square Error) of these labels for the four zero-shot classification methods. The results confirmed that the proposed method can obtain higher accuracy label sets for zero-shot classification.

In the midst of declining population, the Japanese bureaucracy is currently urged to maximize its efficiency. In this regard, the Government of Japan once stated it would begin to study the utilization of generative AI for drafting governmental answers for Diet debate, of which any outcome hasn’t been disclosed yet. While making use of LangChain, gpt-3.5-turbo and Web API of full-text database for the minutes of the Japanese Diet, the author, who was personally in charge of North Korean affairs from January 2003 to March 2005 in the Japanese diplomatic service, compares the generated texts by the drafting system based on LLM and the minutes of the governmental answers to the same questions on diplomatic issues of North Korean affairs delivered in the Diet debates in the same period. The BERTScore turns out to remain below 0.7, while the incapability of the drafting system using the generative AI to express minute nuances needed in governmental answers to questions regarding such issues as North Korea is quite obvious based on human evaluation. This is both partly and technically rooted in typical structures of generative AI, which still limits further improvement of efficiency in the Japanese bureaucracy, particularly diplomatic service.

The COVID-19 pandemic changed our lifestyles. It is expected that the changes in depression state also occurred because of these changes. In this study, we used the questionnaire responses that asked high school students in Tokyo about their depression states before, during, and after the period of the COVID-19 pandemic and analyzed the group characteristics of changes in depression state as a landscape using energy landscape analysis (ELA), a method of multidimensional (time-series) data analysis. As a result, we were able to quantitatively confirm the depression state changes as the energy barriers. We were also able to detect how the energy barrier changed in the COVID-19 pandemic and found that the energy barrier was relatively high in the COVID-19 pandemic. These results are consistent with those of previous studies and suggest that ELA can be used for the psychiatric questionnaires analysis.

A supraglottic airway device (SGA), used for airway management during general anesthesia, provides less hemodynamic change and airway injury than tracheal intubation. However, ventilation can be difficult if laryngospasm occurs when using an SGA. Laryngospasm, an airway reflex triggered by pain or secretions, is more likely in young children and with inexperienced anesthesiologists. To use SGAs safely, it is imperative to maintain airway patency. We aimed to develop a prediction model for ventilatory difficulty in pediatric patients undergoing general anesthesia with an SGA. We analyzed the anesthesia time-series records of the 579 children. The model was trained using the data between 2018 and 2022 and was evaluated using the data from 2023. A multivariate statistical process control model achieved a 57% recall and a 0.65 times/h of false positive rate. In conclusion, we could detect approximately 60% of the ventilatory difficult events during pediatric SGA use.

This study presented multiple prognosis prediction methods for Long QT Syndrome type 2 (LQT2), a hereditary arrhythmic disease. Specifically, we compared the effectiveness of a traditional method using Multiple Sequence Alignment (MSA) with that of a Foundation model (ProtBert) pre-trained on a large dataset without MSA. The results indicated that the method using ProtBert with reconstruction showed the highest prognostic accuracy, suggesting that it is effective in predicting LQT2 prognosis. It is also applicable to the analysis of genetic variants, and this method may be particularly useful for prognosis prediction in situations where annotation costs are high and labeled data sets are scarce.

Acute liver failure (ALF) is a severe condition characterized by rapid deterioration and coagulopathy, which in part evolves from acute liver injury (ALI). Despite its severity, effective treatment for ALF is limited, with liver transplantation being almost the only available therapy. In this study, using 320 patients with ALI at Kyushu University Hospital, we found that prothrombin time activation rate (PT%) is an important indicator of individual ALF status. Furthermore, by adapting unsupervised clustering to the time course patterns of PT% values during the first 7 days after admission, we identified 6 stratified groups with different patterns. Furthermore, by combining a mathematical model with machine learning, we demonstrated that PT% dynamics during the first 7 days after admission can be predicted at the individual level. The model provides important insights for personalized medicine and optimal healthcare resource allocation, as well as new perspectives on the treatment and understanding of ALF.

Machine learning has enabled accurate estimation of true age from fundus images. However, it is unclear which image features are important, and which parts of the image are most clinically relevant, for age estimation. Although methods such as Grad-CAM and DiDA can be used to interpret where the machine learning model looks for inferences, most studies have focused on object detection and classification, with few investigating regression problems. In this paper, we apply DiDA to two different models for age estimation from fundus images and examines the relationship between the sizes of the responding regions and the errors in the estimated age. In addition to counting the number of reacted pixels, we divided the fundus images into three regions, and the counted reacted pixels in each region to decide whether the region was reacted or not. The results show that a smaller error between the actual and estimated age, leads to a greater number of reacted areas and the more accurate the estimated age. These results suggest that the DiDA algorithm can be used to extract the confidence level of the age estimation model.