The CAP Babel Machine

Manually labeling large corpora can be tedious and often demands extensive domain expertise. Recent advancements in machine learning and natural language processing hold the potential for language models to surpass human-level accuracies in text classification tasks.

In our experiments, we fine-tuned Large Language Models (LLMs) across various language and domain corpora using 21 categories from the Comparative Agendas Project codebook. Leveraging multilingual XLM-RoBERTa models, our pipeline generates state-of-the-art outputs for selected language-domain pairs and domains, such as media or parliamentary speech. We fine-tuned a total of 41 models, with the first being a pooled model trained on the entire (multilingual) dataset, followed by 9 models using language-specific datasets, and 10 using domain-specific datasets, while the remaining models were fine-tuned on language-domain datasets.