This research proposal describes the development of a virtual space platform aimed at transforming learning and collaboration in the post-corona era. This research uses technology and the Unity platform to construct a space where learning data can be collected and where students attend lectures that are close to reality.
Abstractive document summarization is one of the most important tasks in natural language processing. Many approaches based on large language models have been proposed. However, it is known that the output of LLM often includes hallucinations, such as factual inconsistency. Therefore, detecting factual inconsistencies in a summary is one important task for summarization. One solution for the detection is to utilize machine learning techniques. In general, machine learning approaches require a large number of training data to generate a robust model. However, it is difficult automatically to collect article-summary pairs with factual inconsistency from the Web because hand-written summaries on the Web are usually correct. Moreover, some existing datasets are written in English. In this paper, we propose some approaches to construct Japanese datasets with factual inconsistency automatically. For this purpose, we utilize two approaches from previous studies: FactCC and SumFC. In addition, we propose a new approach to construct summaries with exaggerated expressions, as a variety of factual inconsistencies. We call the datasets JFactCC, JSumFC, and JExnoS. For JExnoS, we utilize a two-stage approach based on GPT-4 and BART for the generation of summaries with exaggerated expressions from correct article-summary pairs. We also verify the usefulness of each constructed dataset through an experiment about factual inconsistency detection with BERT.
This study introduces an automated menu planning system for schools that leverages machine learning to optimize and streamline menu creation, focusing on nutritional guidelines, cost-effectiveness, and variety. The system comprises two main algorithms: the menu combination algorithm, which employs content-based filtering to select diverse recipes by analyzing and quantifying recipe features like color, flavor, and ingredients; and the menu optimization algorithm, which adjusts ingredient quantities to adhere to budget and nutritional standards. To evaluate the system's effectiveness, satisfaction surveys were conducted among school lunch menu planners, using a seven-point Likert scale to measure feedback on menu variety and cost-effectiveness. Results showed that the system aligns well with nutritional and budgetary requirements, with minimal errors in nutritional content and ingredient costs, and substantially reduces menu planning time.
In this paper, we handle aspect term extraction with In-Context Learning (ICL) as the target task. ICL is a method for learning Large Language Models. Instead of updating the model's parameters, prompts are provided to guide it to perform a target task. While the strength of ICL is that the model does not need to undergo additional training, the prompt includes input-output instances that cause instability in accuracy. The instances are called few-shot. Hence, selecting appropriate instances has the most important role in ICL. For this purpose, we propose a selection method with active learning. Active learning is a method of selecting instances that are useful to the model for training from unlabeled data. We regard the active learning-based approach as a sub-task for the target task. We introduce two types of sub-tasks and evaluate the effectiveness of them in the target task.
Huang et al. (2021) introduced a quantum key agreement scheme using a quantum search algorithm. They initially designed a two-party protocol, later extending it to accommodate three-or N-party scenarios. Notably, Huang et al.'s protocol avoids the need for complex quantum technologies such as entangled quantum particles. However, it is observed that Huang et al.'s protocol imposes a burden on a single participant when applied to multiparty scenarios. This paper contends that such centralization to a single participant is suboptimal for multiparty settings. Thus, our work seeks to build upon the foundations laid by Huang et al. in a manner that addresses this limitation, presenting a modified framework tailored to accommodate multiparty scenarios while distributing computational responsibilities more evenly among participants.
This study aims to explore a state of health (SOH) prediction method and publicly available experimental data from the NASA database. The data is initially filtered using correlation coefficients, and then support vector regression is employed to establish and train a machine learning model. This model can predict the battery's SOH based on operating data such as charging-discharging curves and temperature. Experimental results indicate that using the four features identified by Spearman correlation coefficient for support vector regression yields the best training results, with an average relative error of only 0.0142. This represents an improvement of 11.15% compared to using Pearson correlation coefficient, 172.88% compared to grey relational analysis, and 14.23 % compared to using all features.
In recent years, the collaboration between operators and machines in tasks has gained importance. The scheduling problem by the production method combining operators and machines is called the “Multi-Task Simultaneous Monitoring Dual Resource Constraint Scheduling (MTSSDRCS) problem” and it is crucial to solve it appropriately. An approximate solution for the MTSSDRCS problem has been proposed in previous research because the problem is NP-hard. The decoding scheme is used for the approximate solution. Also, it is necessary to consider the differences in operator ability because they decrease the accuracy of the schedule. However previous studies did not consider these differences. This study solves the MTSSDRCS problem considering the differences in operator ability and proposes decoding schemes.
With the increasing use of technology, the amount of data has been grown rapidly. The data can be reduced by using data deduplication functions. It finds the redundant data stored in a file system, and it deletes the data; hence, the amount of data can be reduced. However, the function affects the performance because of the many read I/O and comparing data. To mitigate the performance penalty, it is effective to narrow down the files processed. Conventional methods use file metadata, such as file size, and a hash value generated from file contents as indicators. However, if the many files are stored in a file system, then the methods cannot narrow files down efficiently. Moreover, if the large files are stored in the file system, then the calculation of hash value becomes a high CPU load. The indicators are recently insufficient. Therefore, a new indicator is necessary for the method. This paper proposes a novel method, called NLPDedup, to narrow files down for data deduplication functions. NLPDedup uses the file names as an indicator for narrowing down target files. This paper describes the overview of NLPDedup, how NLPDedup determines the target files, and the evaluation results. From the evaluation results, the thresholds of indicator need to be set in terms of the algorithms and use cases.
This study proposes an automated checkout system utilizing improved YOLOv5 technology to classify product images effectively. Traditional self-checkout systems often rely on barcode scanning, leading to time-consuming processes and potential errors. Our system eliminates manual barcode scanning by analyzing product images for quick identification and classification, achieving an impressive classification accuracy of 99.1 % in experimental testing. This research contributes to significant improvements in retail automation, enhancing checkout speed, accuracy, and convenience for both customers and retailers.
Most epidemiological analysis of COVID-19 data are aimed to predict the number of cases in the near future, proposing linear or non-linear models. In this study, we aim to uncover the dynamic relation among the most salient pandemic measures published in the well-known Our World in Data site for Japan. Japan has been proposed as a showcase for pandemic management in some periods of the pandemic, hence the discovery of relations among main epidemiological variables should be highly relevant for the assessment of the pandemic management effects. The wavelet coherence analysis (WC) provides a multi-scale correlation analysis with good time and frequency localization. It allows also to analyze the phase of the synchronization of the variables. The main epidemiological outcome considered is the time series of COVID-19 deaths. The analysis finds that this outcome was mostly unaffected by the social control measures at all time scales considered. Also, the vaccination doses and the main outcome are uncorrelated at all time scales to for most dates in the pandemic period. However, at some specific dates they show high in-phase correlation for short periods, which is contrary to the intuitive expectation of the effect of vaccine doses.
This paper aims to demonstrate the existence of prototype networks and their learning process. Prototype learning aims to acquire the simplest form or configuration within the given network resources. This prototype learning is assumed to be naturally integrated into conventional learning, referred to as “natural learning”. Additionally, prototype learning in this natural learning is augmented by separately employing prototype learning, which can be termed “forced learning”. The existence of prototype networks is not easily identified, as they are sometimes weakly activated. To explicitly identify prototype networks, we introduce potentiality and corresponding ratio potentiality. They are analogous to entropy and divergence in conventional information theoretic methods but are developed for interpretation, aiming to clarify more detailed characteristics of learning. The method was preliminarily applied to the bankruptcy dataset, which was small, but it has been challenging to improve and interpret the final results. The results confirmed that the new method could elucidate the existence of prototype networks at the beginning of learning in the natural learning paradigm. This identification of prototype networks was reinforced by employing the forced method. Finally, the new method demonstrated the possibility of discovering new relationships between inputs and outputs beyond linear correlation coefficients.
There are several approaches to the simulation of pedestrian crowds. Previous studies have treated pedestrians as agents or particles in these simulations. Simulations that include random elements must be run many times and use the average result. In this paper, we propose a method for pedestrian simulation by means of a probability value cellular automaton. Pedestrians are represented by a probability distribution and moved according to rules. The congestion distribution of the pedestrian flow can be computed with a single simulation by the calculation of the spatio-temporal probability of pedestrian presence. The validity of the method was verified by experimentation.
The rapid development of AI technologies has made machine interaction a daily norm, such as OpenAI's ChatGPT and Microsoft's Copilot, leading us to develop a personalized book recommender system utilizing Large Language Models (LLMs). By analyzing user behavior, employing natural language processing, and applying a fine-tuned recommender model, this system notably enhances the accuracy of its book suggestions and user satisfaction. Our comprehensive evaluations reveal that this framework significantly surpasses traditional models in delivering personalized content that aligns with users' unique reading preferences and dialogue histories. This research delves into the capabilities of LLMs to offer tailored book recommendations, highlighting the system's ability to synergize user data with book content for improved recommendation precision. It also examines user interactions with LLMs, offering valuable insights for future AI-driven recommender systems.
Management of technology development requires adequate measurements. This paper discusses a broad effort to define, deploy, and adopt software quality measures, especially software reliability in an industrial setting. The industrial projects described here began before the emergence of the Web and continue to the current day. The approach and results of these quality metrics support the technical management of a variety of large-scale software platforms, their operations, and the improvement of quality across numerous iterations of development. Additionally, a comparison of classical customer oriented and dynamic software metrics with those commonly applied within Agile and DevOps environments is discussed in the context of the evolution of technology management and measurements. These metrics include Software Reliability, Test Effectiveness, and Code Coverage. Management decisioning using these metrics is traced through several real-world industrial projects across differing companies. Conclusions are linked to current trends in managing Software Development such as ITIL, Agile methods, and DevOps.
Trust is dynamically evolving from human-human interactions to human-automation/artificial intelligence (AI) engagements. Previous research has characterized trust as elusive. This study posits a society wherein trust is readily captured and utilized within a system, focusing on the interrelations between trust, reliance, and their components concerning technology. Although prior studies have identified the trustor and trustee as essential elements in fostering trust, this research introduces an architecture that incorporates a model encompassing context and proxy trust on the part of the trustee. To address challenges highlighted in previous studies, we have developed a model demonstrating that observation between the trustor and trustee is crucial, and that design features, system performance, and dependability correlate with trust, thereby facilitating proxy trust. The explicit illustration of trust dynamics between humans and between humans and technology elucidates the relationship between trust and technology, an aspect that will gain increasing significance in the future.
Cyber defense exercise is effective in training security personnel against the sophisticated cyber-attacks. In this exercise, beginners operate a virtual service and defend the service from attacks. A student can exercise only one defense method for one attack in exercise, so he/she cannot judge whether the defense method was appropriate. It is thought that this problem is able to be solved by trying defense methods many times through Trial-and-Error during exercises. We develop a defense training system with a Trial-and-Error function by using the snapshot function of virtual machine. In order to conduct cyber defense exercises, a cyber Attack function and Score measurement function are required. In this paper, design and implementation of these functions are described. Furthermore, evaluation of them using the same and different methods is also described. The evaluation results show that the cyber Attack function and Score measurement function are effective.
In this study, we explore the Management Discussion and Analysis (MD&A) section in the Japanese Annual Securities Reports, a mandatory disclosure known for providing crucial qualitative information about management perspectives. Our research primarily utilizes ChatGPT to extract qualitative expressions within the MD&A texts. Firstly, we quantify the extent of qualitative information presented in their MD&A and compare the proportion of qualitative expressions across different companies. Our hypothesis that a higher prevalence of qualitative information may indicate a deeper understanding by management of their company's business model, market environment, and strategy, potentially leading to more accurate performance forecasts. We then analyze the impact of the proportion of qualitative expressions in the MD&A on the accuracy of management earnings guidance in the financial results summary. We aim to understand how the nature of information in MD&A -whether more qualitative - correlates with the precision of managerial predictions on company performance. We have uncovered a statistically significant negative correlation between the quantity of qualitative expressions and the magnitude of management forecast errors. This finding implies that when MD&A documents contain more qualitative insights, the predictions made by management about future financial performance tend to be more accurate.
The widespread of the Internet has realized the electronic commerce (EC) that enables us to make purchases online and is used by a lot of people throughout the world. However, with the increasing numbers of users and products, the mismatch between users and products are getting prominent. In order to avoid such troubles, on EC sites, users can post a review on the commodity they bought. Then other users can grade the review as helpful. These functions help users refer to the review of the item and consider whether to buy it. Here, there could be a method to solve more mismatches between users and products by using the impression from the statements of those reviews. Therefore, in this paper, with the aim of preliminary experiment, we try to experimentally extract the impression from the review contents. We conduct impression evaluation experiment on assessing twenty low-rated reviews using impression words. As a result of factors analysis applied to the experimental results, three factors representing the style or content of the statements are obtained. This result could imply more useful outcomes with future additional experiments.
