Translating academic texts containing specialized terms not included in the training data of Large Language Models (LLM) presents a significant challenge. We introduce a new approach that utilizes a Retrieval-Augmented Generation (RAG) model with a Japanese-English dictionary and two knowledge graphs constructed from Japanese and English datasets. Experiments with RAGs have demonstrated that dictionary-based RAGs assist in translation. We also explore a RAG approach that leverages two knowledge graphs based on Japanese and English data, revealing that its performance is not inferior. Moreover, we suggest that with an appropriate method of constructing knowledge graphs, this approach could potentially improve translation accuracy.

Some interoperable metadata terms have different roles from the labels of the terms themselves, depending on the domain in which the terms are used. However, it is not enough of vocabulary definition and relationship definition between terms, so it is difficult for metadata schema designers have less knowledge about metadata to find interoperable terms from the domain-specific words. We propose a method that maps interoperable metadata terms to thesaurus concepts that express the roles of the terms in the specific domain using LLM. We thought that we can link the metadata terms to the words that represent role of the terms by combining interoperable terms with domains that the terms are used in. Even without knowledge of the metadata, it is promising that the designers will find more interoperable terms by searching for the terms from the designers' words via the thesaurus.

Although various knowledge graph embedding models have been proposed, they cannot enhance prediction of subsumption and membership relations from ontologies. Therefore, more attention has been paid to ontology embedding methods that uses lexical information or logical structures of ontologies. However, most of the methods do not employ lexical information together with logical structures. In this paper, we propose a method capturing both lexical and logical characteristics from ontologies. In this method, we extend the complex embedding of RotatE by the hierarchical radial coordinate of HAKE and initialize entity embeddings by the vectors learned from lexical information. In the experiments, we show that the proposed method outperforms existing methods for the subsumption prediction tasks on the benchmark ontologies FoodOn, GO, and HeLiS.

The purpose of this study is to construct Linked Open Data that contributes to the inheritance and understanding of local foods.Local foods contain important information on regional characteristics and culture, but due to changes in modern dietary habits, opportunities to pass on local foods at home are decreasing. Therefore, we conducted a survey on local foods and found that the most common reason for wanting to make local foods was "wanting to pass them on," while the most common reasons for not wanting to make local foods were "time-consuming, troublesome, and labor-intensive" and "it is difficult to prepare the ingredients. Based on these results, we hypothesized that simplifying existing recipes using a large-scale language model and converting them to LOD would contribute to the transmission of Local food, and created a simplification system for Local food recipes.In this study, we attempted to improve the accuracy of the recipes by utilizing human "modification comments" on the generated simplified recipes. We also examined how to effectively utilize these comments.

Recent multimodal platforms consist of modals centered on voice, text, images, and music. Text-to-image is often used, such as in anime character generation, and its quality is comparable to that of creators, and it is becoming an alternative to AI creators. Furthermore, Image-to-Video is also emerging. These are based on Text, and are gaining social acceptance. On the other hand, there are few attempts between Image-to-Music and Music-to-Image modals. Technically speaking, it can be thought of as a method that individually tokenizes multiple different data such as audio, text, images, and music, and performs multimodal understanding and generation in an autoregressive manner as a large-scale language model (LLM). One of the reasons why it becomes a black box is that it is disconnected from human senses, and it is important to understand it from the perspective of knowledge graphs and ontology. In this paper, we consider the interpretability of Image-to-Video and Image-to-Music, and provide future prospects.

Efficient extraction and organization of knowledge graphs from encyclopedic documents, such as Wikipedia, can be expected using natural language processing tools.However, there are differences in the target domain of the knowledge description, and it is not easy to adjust the algorithm for knowledge graph extraction according to the target domain. This paper proposes a general-purpose natural language processing interface for knowledge extraction that can be customized to meet specific requirements and characteristics according to the target domain of the knowledge description.

While the use of generative AI is attracting considerable attention, employing large-scale language models (LLMs) has emerged as one of the most significant issues in the field of knowledge engineering. In the application of LLMs, the ability to explain the generated content represents a critical challenge. This challenge is shared with the Knowledge Graph Reasoning Challenge, on which the authors have been focusing. Against this backdrop, we have designated the construction of knowledge graphs using LLMs as the task for the Knowledge Graph Reasoning Challenge 2023. This paper offers an overview of the event and its outcomes.