Extracting Organization-Location Relations from German News Articles: A Comparative Study of Techniques

CRIF S.p.A., a leading organization in the field of business information extraction, is currently developing a project to extract information about companies from German news articles that may not be present in existing business databases. A key aspect of this work involves extracting the location of public or private companies, which is crucial for disambiguating companies with similar names, addressing incomplete company names in article texts, and distinguishing between different headquarters. To achieve this, several relationship extraction techniques have been explored to evaluate their performance. The findings align with recent research, confirming that prompting large language models, specifically GPT-3.5 Turbo, is the most effective approach. However, even more lightweight approaches, like question answering, also show significant promise and merit further investigation.