NKOS Workshop

Large Language Models (LLMs), such as ChatGPT, are increasingly being integrated into healthcare, and their application in mental health, particularly in managing depression, presents both potential benefits and challenges. This study investigates how LLM-based chatbots can empower users by providing instant, personalized support while addressing the need for robust safety mechanisms in sensitive mental health contexts. Participants will engage in one-hour remote interviews, interacting with a ChatGPT API-powered chatbot focused on the self-management of depression. Reflexive thematic analysis will be used to identify themes related to user perceptions and potential harms. Anticipated outcomes include insights into the effectiveness of chatbots in managing depression, potential harms, and design implications for safe and effective LLMs for depression self- management. The findings aim to enhance knowledge organization systems within LLMs by improving the structuring and access of mental health information. Preliminary results will be presented at the workshop, showcasing the data collection webpage developed using the ChatGPT API.

This demo proposal presents a case study that uses Gene Ontology (GO) and ML/AI algorithms to design and create KO-derived datasets for ML/AI applications. We discuss the characteristics and requirements of KO practices and products in implementing ML algorithms. The focus of this demo is on how knowledge organization systems can be utilized to derive datasets that can deliver quality and trustworthiness for achieving the precision, computing of semantic similarity, and interoperability in these algorithms.

As with many areas of research and practice, the cultural heritage domain has shown increasing interest in the use of AI in recent years, with cultural heritage institutions such as libraries, archives, and museums actively exploring the use of AI tools in their workflows. Large language model (LLM)-based text applications including ChatGPT have been touted as holding great promise for cultural heritage work. One of the most challenging parts of producing library metadata specifically may be subject cataloging: the assignment of subject headings and classification numbers. This task requires cataloger fluency in the formal and often complex knowledge organization systems (KOS) used to represent aboutness and genre in bibliographic records. The work presented here is part of a larger, ongoing research project assessing the effectiveness of AI tools for performing subject analysis and representation tasks for cultural heritage data. In this presentation, researchers offer the results of a structured test of freely available AI tools to assign headings from Library of Congress Subject Headings (LCSH) to library materials. The findings add further empirical evidence into current discussions concerning the quality and reliability of AI-performed metadata work, and, more broadly, contribute to the growing discourse around the use of AI in applying KOS.

Patent citation relationships constitute a citation network, and the predictability of edges in a network is a frontier research issue in complex networks. This article explores the prediction model of patent citation relationships. By integrating patent technical text content and classification code features, a graph neural network is trained for patent citation link prediction. These aim to provide methodology support for technology knowledge diffusion and patent data management. This study collects patent data in the field of quantum sensing, constructs a network based on patent citation relationships, and extracts text features such as technical problems, solutions, functions, and effects. This article proposes a new link prediction model framework based on graph neural networks, taking into account the characteristics of natural language in patent documents. Addressing the characteristics of natural language in patent literature, this article proposes a new model framework for link prediction based on graph neural networks.

In terms of model framework, we initially employ the GraphSAGE model on the training citation network to obtain the embedding vectors of patent nodes. Then, the semantic vectors of patent technical text are derived by pre-trained models such as PatentBERT. These two sets of vectors are then integrated and fed in a Random Forest model. Ultimately, we derive the predicted probability values for patent citation link prediction. Furthermore, in terms of interpretability, this study constructs a decision tree model based on the integrated results of the two sets of vectors. This model effectively measures the impact of multidimensional technical text content, local network structure, individual heterogeneity, and other factors on network edge formation.

Generative artificial intelligence (GAI), particularly those based on large language models (LLMs), has become an increasingly important tool in digital humanities. It enhances research efficiency in tasks such as content analysis, keyword extraction, automated metadata creation, and data management, uncovering previously difficult-to-observe phenomena and tackling challenging issues. Beyond data generation, GAI’s rapid content analysis and knowledge structure design capabilities offer new exploratory directions for constructing and designing thesauri based on Knowledge Organization Systems (KOS). Using the multilingual "Art & Architecture Thesaurus" (AAT) developed by the Getty Research Institute (GRI) as an example, the Academia Sinica Center for Digital Cultures (ASCDC) has collaborated with GRI for over a decade to address the inadequacies of localized cultural vocabulary. The Chinese language and concepts of material culture are converted into English and integrated into the AAT through translation and mapping. During this process, the conceptual structure of controlled vocabularies in Chinese and English terms presents multiple alignment patterns, and a systematic methodology has been developed to support editorial work. This study aims to explore how GAI can assist in constructing a structured thesaurus based on the cultural conceptual vocabulary related to the Confucius Ceremony, with the goal of contributing this localized vocabulary to AAT.

The development of multimodal large language models(MLLMs) provides new solutions for knowledge organization and representation of multi-modal cultural memory resources. However, for the Knowledge Organization and Representation of some special cultural memory resources such as text, images, audio, and video resources related to the Guqin Subtractive Character Notation(see fig.1 and fig.2), the existing MLLMs need further optimization to achieve the expected results. Guqin Subtractive Character Notation is a distinct notation system rich in Chinese cultural significance, differing from both simplified notation and traditional staff notation.It shows the fingering techniques for Playing Guqin.It is not a Chinese character and can be recognized only by a very small number of professionals who have undergone long-term training.It cannot be recorgnized with the existing OCR technologies.
This study will use multi-modal Guqin Subtractive Character Notation resources as training data(see tab.1), and combine with Guqin ontology application profile and RDF data as prompt tuning data to explore a vertical application path of a MLLMs in the field of cultural heritage, and develop a prototype system to display the research results.The screen recording presented preliminary research results. By using the multimodal resources and Guqin ontology with RDF data as instruction fine-tuning data to fine tune the multimodal large language model, the cross modal retrieval with images and audio as input query can be achieved. The ultimate goal of this study is to use the optimized MLLMs to help more people understand the Guqin Subtractive Character Notation, especially those in the large collection of ancient books in the libraries.

This paper presents work on using Large Language Models (LLM) for disambiguating Named Entity Linking candidates, which is meant for enriching the metadata of textual documents by linking them to Knowledge Organization Systems, a.k.a domain ontologies, and Knowledge Graphs. We propose a zero-shot classification method that has similarities with Retrieval-Augmented Generation (RAG), and discuss an under-development prototype tool that allows for human intervention when making final disambiguation decisions, especially when this cannot be reliably carried out in automatic fashion. The focus of this work is on Finnish texts, so our methods must take into account the particularities of this language and the resources available for processing it.

Artificial Intelligence (AI) is ubiquitous, shaping various facets of our global society. Understanding its impact and application across different domains is essential for members of the general public and community stakeholders in government, education, and industry. This demonstration will showcase a dashboard that leverages AI technologies and Knowledge Organization Systems (KOS) to track and analyze AI’s presence and impact worldwide. In this demo, I discuss how AI is applied in government, education, and industry, using the AI dashboard as a guide. I discuss what KOS we’ve used to create and power the dashboard, which is a tool to track AI’s ubiquity. Although still a work in progress, I will discuss how this dashboard uses web scraping tools to collect real-time data from sources such as news websites, government reports, academic publications, and industry reports, demonstrate how APIs are used to gather structured data, and the use of machine learning to normalize data from these sources to ensure consistency and accuracy through deduplication, standardization and validation of the data collected. Additionally, I will discuss some of the issues we've faced with copyright, restrictions, and limitations of API usage, and how we are dealing with bias mitigation. I will also talk about how ontologies define the relationships between AI and its subdomains, the technologies it powers, applications, and sectors. These ontologies provide a structured framework for organizing the data collected. The hierarchical classification within the ontologies makes it easier for users to navigate and understand the data, offering a clear view of the AI landscape. Controlled vocabularies ensure consistency in the terms used to describe AI technologies and applications across different data sources. This standardization enhances data integration and retrieval. A thesaurus captures synonyms and related terms to improve search capabilities within the dashboard. Taxonomies categorize AI and its impact, which allows users to filter and explore information based on specific criteria such as sector, region, and technology type.

Conferences provide a way to learn the most recent developments in a wide range of subject areas.  In particularly fast moving areas like cancer research or imaging and photonics there are literally 10’s of thousand papers submitted each year for consideration. The number of attendees can reach 10’s of thousands as well.  Every six to twelve months.  There was a strong need for automation.  The process involves peer review, categorizing into conference tracks and then matching the attendee to the sessions of most interest to them.  Using a custom topical taxonomy we demonstrate how the papers are channeled to approbate reviewers, then conference tracks based on top terms, and then suggested to the attendees based on their individual semantic profiles.  The two case studies show the process as well as the resulting mobile apps for use in the conference experience.