Fast and Accurate Quotation Attribution in Literary Texts

Quotation attribution—the task of identifying which character speaks a given line of text—is essential for computational literary analysis. While recent large language model (LLM) approaches have achieved high accuracy, they are computationally expensive and impractical for large-scale studies. Conversely, traditional methods that attribute quotations independently are efficient but often lack the necessary precision. This paper addresses this trade-off by proposing a new, efficient formulation called joint scoring.

Inspired by span-based coreference resolution, the authors replace the standard "direct scoring" approach (where each quotation is processed in isolation) with "joint scoring." In this framework, multiple quotation attribution decisions within a large, shared context window are resolved simultaneously. By fine-tuning a ModernBERT encoder to optimize these decisions jointly, the model captures local conversational patterns and coreference structures more effectively. The authors also release a modified version of the popular BookNLP pipeline, incorporating this system to facilitate broader adoption.

• State-of-the-Art Performance: The joint scoring model achieves 94.5% accuracy on the Project Dialogism Novel Corpus (PDNC), outperforming previous LLM-based benchmarks.
• Efficiency Gains: The system is 20 times faster than standard direct scoring methods and over 1,000 times faster than LLM-based approaches on an A100 GPU.
• Mechanism of Success: Analysis shows that joint scoring excels because it better preserves the long-range anaphora resolution signals already present in pretrained encoders, which are crucial for linking pronouns to the correct character.
• Coreference Importance: Contrary to some recent trends, the authors find that including pronominal and nominal mentions (via coreference clusters) significantly improves accuracy compared to using only named mentions.