TL;DR
This paper tackles the challenge of coreference resolution in literary texts, a task crucial for computational literary studies but underserved by existing NLP tools trained on news or dialogue. The authors analyze unique literary phenomena, propose adaptations to annotation guidelines, and introduce a new tool to facilitate the annotation of long, complex literary works.
What Problem Does This Paper Address?
Coreference resolution, the task of identifying all mentions that refer to the same entity in a text, is a fundamental NLP problem. While significant progress has been made, most automatic coreference resolvers are trained on news articles, broadcast conversations, or web text. However, literary texts present a distinct set of challenges. Their purpose is not solely information transfer but also poetic description and storytelling, leading to sophisticated language, rich vocabulary, and complex syntactic structures. Existing annotation guidelines and tools, often developed for more straightforward text types, struggle to adequately capture the nuances of literary coreference. This gap hinders the application of NLP techniques to analyze literary works, such as character depiction, plot analysis, and sentiment analysis.
Methods
The authors' approach is two-fold: analysis and tool development.
- Analysis of Literary Phenomena: They conducted an annotation study on a corpus of thirteen literary texts from different centuries and genres. Through an iterative process of annotation and discussion, they identified and analyzed phenomena specific to literary texts that impact coreference resolution. This involved refining existing annotation guidelines based on these observations.
- Annotation Tool Development: To address the challenges of annotating long literary texts with many discourse entities, they developed a new annotation tool. This tool is designed for fast annotation, supports keyboard-driven interaction, and conceptualizes coreference as equivalence sets rather than binary links. It also allows for the annotation of mentions within tokens and handles long dependencies spanning entire books.
Datasets
The paper describes an annotation study conducted on a corpus of thirteen literary texts. These texts span different centuries and literary forms, including:
- Johann Wolfgang von Goethe's The Sorrows of Young Werther
- Leo Perutz's From nine to nine
- Novels Der Mond lacht and Nur ein Druck auf den Knopf
- Plays Miss Sara Sampson and The Robbers
- Six Brothers Grimm fairytales (Frog Prince, Hansel and Gretel, Cinderella, The Town Musicians of Bremen, Rumpelstiltskin, and Rapunzel)
- One English text: A Narrative of the Captivity and Restoration of Mrs. Mary Rowlandson
While the paper doesn't present a formal evaluation of a trained model on a benchmark dataset, it details the process of creating an annotated corpus for literary texts, which can serve as a new benchmark.
Key Insights and Ideas
- Domain Matters Profoundly: The study strongly emphasizes that coreference resolution is highly domain-dependent. Literary texts, with their unique linguistic and narrative structures, require specialized approaches and guidelines compared to news or dialogue.
- Genericity is Tricky: Literary texts frequently switch between generic and non-generic uses of noun phrases, making it challenging to distinguish between class references and individual references.
- Narrative Spheres and Knowledge Layers: Literary texts operate on multiple "spheres" (narrator, narrated) and involve complex layers of knowledge for the reader, narrator, and characters. This can lead to intentional ambiguity and requires annotators to adopt a consistent perspective (e.g., reader's point of view after one read-through).
- Entity Development and Ambiguity: Characters and their roles can evolve significantly in literary texts, raising questions about whether referents remain the same. The paper also highlights the intentional use of ambiguity in literature, which challenges traditional NLP approaches that assume a single "ground truth."
- Text Length is a Major Hurdle: Literary works can be hundreds of pages long, making it difficult for annotators to recall entities from early chapters when annotating later ones. This necessitates tools that can manage long-range dependencies and provide efficient search capabilities.
- World Knowledge is Essential: Historical literary texts often assume a level of world knowledge that contemporary readers may lack. Annotators need guidance on how to handle this, such as looking up relevant information to infer intended references.
Key Results
The paper's primary contribution is not a set of quantitative performance metrics but rather a theoretical and empirically-grounded analysis of literary coreference phenomena and a proposed framework for their annotation.
- Identification of Literary-Specific Phenomena: The authors detail several phenomena that differentiate literary coreference from other domains, including complex entity development, multiple narrative spheres, genericity shifts, and the impact of world and text knowledge.
- Proposed Guideline Adaptations: They outline specific modifications to existing annotation guidelines (based on NoSta-D) to accommodate these literary peculiarities, such as introducing generic entity types and focusing on entity clusters rather than binary links.
- New Annotation Tool: The development of a specialized annotation tool optimized for long texts and efficient coreference annotation, supporting features like sub-token annotation and equivalence set representation.
Limitations and Future Work
- Limitations:
- The paper acknowledges that making a majority-based decision for intentionally ambiguous annotations might not fully capture the complexities of literary language.
- Quantitative evaluation of the proposed guidelines and tool is not presented, as the focus is on the analysis and adaptation process.
- The distinction between coreference and bridging can be tricky in literary texts, and the resolution of individual cases is left to annotators.
- Future Work:
- The authors plan to use the developed annotation guidelines to create a corpus of selected German literary texts, which will be made openly available.
- They are exploring ways to represent conflicting or diverging annotations within their tool.
- The annotation tool is designed to be easily localized to other languages, suggesting future multilingual applications.
- The paper raises conceptual questions about representing multiple, conflicting annotations in file formats and how to properly evaluate such data, indicating a need for new methodologies in computational linguistics.
Summary updated