The main goal of equipment maintenance management is to keep equipment in optimal working order. Regular equipment maintenance is the key to preventing catastrophic equipment failure. It can pay off in the form of increased operational efficiency and reduce cost. Equipment maintenance service requires a wide area of knowledge, spanning from physicochemical phenomena that cause failure to the correct procedure of actual maintenance operations. This paper presents an ontological knowledge modeling framework to describe, share and inherit the knowledge. As an example of knowledge models described by the framework, functional decomposition trees of piping system maintenance are shown. The knowledge models can be used in various situations and forms, such as education of staff and quick reference for maintenance service.
This article reports on the technology for managing assets and data in smart cities. We adopted Next Generation Service Interfaces _ Linked Data ontology as the management technology. We introduced rules to construct ontology models for the purpose of retrieving information accurately. Finally, we compared searches from ontology and the tree model, and describe problems and countermeasures.
Consistency in category classification is essential for federated search. Consistent categorization through LLM could serve as a potential solution. RAG is utilized to enable LLM to process information based on external knowledge. However, when the information source for classification is a KG, it is not clear what adjustments are effective. In this study, we attempt RAG utilizing a KG as the information source and aim to reveal its effectiveness and limitations.
To support the acquisition of recovery concepts, we have developed a concept learning support system using a Knowledge Graph (KG) and the Large Language Model (LLM). With the increasing popularity of recovery concepts, there is a need for proper conceptual understanding, but the problem is that the quality of instructors is not guaranteed, resulting in misunderstanding of the concepts. In this study, we created a KG of the recovery concept and aimed to support the acquisition of appropriate concepts by providing learning support through an LLM-based chatbot that uses this information. Experimental results with university students suggest that KGs are useful for facilitating the understanding of concepts with clear definitions. On the other hand, in the case of concepts with a wide range of applications, it was suggested that the scale of KG needs to be expanded and its use needs to be improved.
In the automotive industry, competition in the electrification sector which is typified by electric vehicles, is intensifying.To stand out in this field, our company aims to establish a development process that balances reduced development periods with enhanced product quality.For that purpose, we are building a system that interprets a wide variety of in-house product development knowledge through ontologies and leverages it for design support.This paper outlines the design support system and the ontology structure that defines internal knowledge and shares it efficiently across multiple products, also explains our knowledge extraction methodology, which is designed to streamline the construction of ontologies.The method consists of four steps: named entity recognition, relation extraction, zero anaphora resolution, and entity linking, proposing a rule-based method that utilizes the ontology structure for relation extraction and entity linking.Evaluations using in-house documents have yielded good results at each step, notably showing that rule-based methods surpass machine learning methods in accuracy. This demonstrates that high-precision knowledge extraction is feasible within in-house domains by constructing ontologies.Furthermore, the study provided insights into the selection of in-house documents during the initial and middle-to-late stages of ontology construction, as well as the number of documents that should be manually annotated, in terms of the volume of knowledge that can be extracted and the accuracy of named entity recognition.
In the field of life sciences, numerous databases are publicly available in the form of graph structures called RDF (Resource Description Framework). Each node in these graphs is linked to a unique URI representing entities such as genes and proteins. Consequently, these databases contain vast amounts of graph-structured data with numerous URIs. By establishing links to URIs in other databases, these databases become interconnected through links, ultimately forming a massive knowledge graph. As a result, it is essential to develop a foundational infrastructure for storing and searching a significant number of URIs from life science data in this vast knowledge graph, utilizing database IDs and keywords. Presently, Front Coding is commonly employed for URI compression, but it suffers from a limitation in its inability to perform prefix matching searches. Trie is frequently used as an alternative approach capable of retrieving prefixes. This study assesses the compression ratio and search efficiency by comparing the utilization of Front Coding, Trie, and our proposed methods. The evaluation is conducted using a dataset of URIs extracted from life-science RDF datasets.
The dataset provided by the Knowledge Graph Reasoning Challenge for Social Issues (KGRC4SI) is an event-centric knowledge graph representation of avatar behavior in a virtual space. Therefore, it contains information on the entire virtual space, including parts of the virtual space that are hidden in the video. However, some information cannot be obtained by any means, for example, when sensor data from the real world is collected. In this paper, we report a method for predicting the hidden parts using LLM and the results of experiments using this method, using the KGRC4SI dataset as complete data for training.
Local food contains important information on regional characteristics and culture, but opportunities to pass on Local food are decreasing due to changes in modern dietary habits. Therefore, this study aims to construct Linked Open Data that contributes to the inheritance of Local food. However, it was not clear what kind of data would contribute to the inheritance of Local food. Therefore, we conducted a survey on how people would like to cook their Local food. The most common reason for wanting to cook Local food was "to pass it on," while the most common reasons for not wanting to cook Local food were "labor-intensive, troublesome, time-consuming," and "it is difficult to prepare the ingredients. Based on the survey results, we hypothesized that simplifying existing Local food recipes and making them easier to prepare using a large-scale language model would contribute to the transmission of Local food. Based on this hypothesis, a recipe simplification system using a large-scale language model of Local food was developed. However, the recipes actually generated were not highly accurate in terms of simplification and regional characteristics. Therefore, by introducing "ingredient substitution knowledge data," we improved the accuracy of the generated recipes, especially the accuracy of ingredient substitution, which was inaccurate.