Lee Kwan-kyu (left), President of the National Institute of Korean Language, poses with Francesco Maria Tucci, CEO of Babelscape, and Roberto Navigli, the company’s head of research, during a MOU signing ceremony held at the institute in Seoul on Sept. 16. Courtesy of the National Institute of Korean Language
Babelscape is collaborating with the National Institute of Korean Language (NIKL) and LiST to bring Urimalsaem to BabelNet. Urimalsaem is a comprehensive, openly accessible dictionary of the Korean language, offering a rich source of lexical knowledge to strengthen BabelNet’s coverage of Korean. Working with an open linguistic resource also aligns with our commitment to open research and multilingual knowledge sharing.
This collaboration was formalized through an agreement signed by NIKL and Babelscape, aimed at advancing the development and use of AI-based language resources. BabelNet, the multilingual semantic network developed by Babelscape, now spans roughly 600 languages and draws on resources such as WordNet, Wikipedia, and Wikidata, making it widely used in natural language processing and AI research. The signing was attended by Roberto Navigli, professor at Sapienza University of Rome and co-founder of Babelscape, who created BabelNet, alongside Babelscape CEO Francesco Maria Tucci. NIKL President Lee Kwan-kyu noted that the partnership would help expand the use of Korean-language resources within the global AI ecosystem.
Mapping means taking a word sense from a dictionary and connecting it to a BabelNet synset: a concept that brings together words with the same meaning across languages. For example, the Korean word 개 (gae), in its sense meaning the animal “dog,” can be linked to the BabelNet synset for the domestic dog, which brings together words such as “dog” in English and “cane” in Italian. These connections would make Urimalsaem’s lexical knowledge accessible within BabelNet’s multilingual network.
For this mapping project, we are exploring how Large Language Models can help identify corresponding meanings across the two resources. Building on our experience with dictionary mapping, including LLM-based approaches for European languages, we are adapting and refining these methods for Korean. This work also connects with our research into how LLMs understand and distinguish word senses, explored in Meconi et al.’s Do Large Language Models Understand Word Senses?.
The approach we are developing uses translation to support the search for possible matches in BabelNet, followed by LLM-based comparison to identify the concept that best matches each dictionary sense. We envision a human-in-the-loop framework in which NIKL expert reviewers contribute their knowledge of the Korean language to validate and refine the proposed mappings. Combining automated methods with linguistic expertise will be central to assessing the quality of the results.
This collaboration is an opportunity to expand Babelscape’s ties with the international lexicographic community and enrich BabelNet’s representation of the Korean language. It also gives us a concrete setting in which to investigate how LLMs can support the integration of lexical resources, helping connect linguistic knowledge across languages.