This paper explores applying machine learning to automatic thematic indexing, using two sub-corpora from the Complete Works of Voltaire as a test case. The task is framed as a multi-label classification problem where models assign index entries to text pages.
- The study compares encoder-based models with classification heads against generative large language models (LLMs) fine-tuned via Low-Rank Adaptation (LoRA).
- Model sizes range from approximately 3 to 120 billion parameters.
- The best-performing model, a Mistral family model in 4-bit quantised configuration, achieves F1 scores of up to 0.67.
- The authors argue these figures represent lower bounds due to the inherent subjectivity of professional indexing.
The findings have implications for providing structured thematic access to large-scale literary and historical corpora.