Leveraging Open Large Language Models for Multilingual Policy Topic Classification: The Babel Machine Approach

The article presents an open-source and freely available natural language processing system for comparative policy studies. The CAP Babel Machine allows for the automated classification of input files based on the 21 major policy topics of the codebook of the Comparative Agendas Project (CAP). By using multilingual XLM-RoBERTa large language models, the pipeline can produce state-of-the-art level outputs for selected pairs of languages and domains (such as media or parliamentary speech). For 24 cases out of 41, the weighted macro F1 of our language-domain models surpassed 0.75 (and, for 6 language-domain pairs, 0.90). Besides macro F1, for most major topic categories, the distribution of micro F1 scores is also centered around 0.75. These results show that the CAP Babel machine is a viable alternative for human coding in terms of validity at less cost and higher reliability. The proposed research design also has significant possibilities for scaling in terms of leveraging new models, covering new languages, and adding new datasets for fine-tuning. Based on our tests on manifesto data, a different policy classification scheme, we argue that model-pipeline frameworks such as the Babel Machine can, over time, potentially replace double-blind human coding for a multitude of comparative classification problems.