A Synergic Solution for the Regional Newsrooms: Identifying Regionally Relevant Information for News Stories with LLMs and Knowledge Graphs
Reshmi Gopalakrishna Pillai, Antske Fokkens, Wouter van Atteveldt
Vrije Universiteit Amsterdam
Over the last few years, Large Language Models (LLMs) have become increasingly pervasive and relevant in research involving text analysis. Despite its potential, the critics of LLMs also have rightly raised concerns about the range of its challenges, perhaps the most well-known of which is its tendency to ‘hallucinate’, i.e. to produce factually incorrect or completely false results. Knowledge graph is a structured representation of entities, relations and attributes, which allows for implicit knowledge retrieval based on logical reasoning and is not limited to explicit knowledge and relations. With its ability to handle heterogenous and dynamic data, it has been explored and implemented to support and enhance information from multiple sources in domains such as journalism. However, the creation of knowledge graphs conventionally required human intervention in defining ontologies. The synergy of knowledge graphs and LLMs has emerged recently as a promising solution to these various challenges. Our research, which is currently a work in progress, explores the adoption of LLMs and knowledge graphs in addressing the challenges of regional news media in the Netherlands. Based on a design thinking process, we identify efficient and reliable processing of the incoming information from multiple sources to be a primary challenge in news media, especially those with a regional scope. LLM-powered knowledge graphs can support news media by creating a structured knowledge representation of the regionally relevant entities and their relations. We establish the feasibility of this synergic solution in the specific use case of extraction of relevant background information for a given news event. Using Graph-based Retrieval Augmented Generation, we retrieve and present the relevant information and indirect and possibly interesting relationships between the entities, present in the news event.