TL;DR
This paper investigates what makes a novel "canonical" by analyzing Danish novels from the Modern Breakthrough era. It reveals that while textual features are important, external factors like publisher and price play a crucial role in a novel's journey to literary canonization, challenging purely text-centric views.
What Problem Does This Paper Address?
Literary scholars have long debated what elevates a novel to canonical status. Recent research suggests that canonical works possess unique textual profiles, often linked to innovation and cognitive complexity. However, this paper highlights a puzzle: some non-canonical novels share these same textual characteristics. This suggests that factors beyond the text itself – what the authors call "text-extrinsic factors" – are critical in determining a novel's lasting impact and inclusion in the literary canon. The authors aim to disentangle the interplay between these internal (textual) and external (socio-economic, institutional) influences.
Methods
The researchers employed a multi-faceted approach:
- Document Representation: They used the
m-e5-large-instructlanguage model to create semantic embeddings for each novel. These embeddings represent the novels' textual content in a numerical format. - Text-Extrinsic Feature Extraction: They gathered data on each novel's first edition price, publisher, and author's nationality.
- Classification Task: A Random Forest model was used to classify novels into "Canon" or "Other" categories. This model was chosen for its robustness with mixed data types and ability to handle outliers.
- Experimental Design: They ran classification experiments using various combinations of features:
- Text-intrinsic features (embeddings) alone.
- Text-extrinsic features (price, publisher, nationality) alone.
- Combinations of text-intrinsic and text-extrinsic features.
- Robustness and Analysis: To ensure reliable results, they used downsampling for imbalanced classes and repeated experiments 50 times, averaging the outcomes. A key part of their analysis involved a deep dive into "false positives" – non-canonical novels that the model incorrectly classified as canonical, to understand why they might have been overlooked.
Datasets
The study utilized a corpus of 838 original Danish and Norwegian novels published between 1870 and 1900, known as the "Modern Breakthrough" era. This corpus was accompanied by metadata including book price, publisher, and author nationality. The categorization of novels into "Canon" or "Other" was based on previous work by Feldkamp et al. (2024b), which drew from official Danish canons and expert-based literary entries.
Key Insights and Ideas
- Beyond the Text: The study strongly suggests that textual similarity alone is insufficient for a novel to achieve canonical status. External factors are significant players.
- Publisher Power: Publishing houses, particularly larger ones, appear to have a substantial influence on canonization. Even novels with canonical textual profiles might be overlooked if not supported by the right publisher.
- Price Matters: Book prices also seem to correlate with canonical status, potentially reflecting accessibility and market positioning.
- Nationality's Nuance: While author nationality showed some predictive power, its role was more complex and sometimes less decisive than publisher or price.
- "Non-Canonical Canonicals": The analysis of false positives highlights the existence of novels with canonical textual features that, for reasons outside their text, failed to break through. This points to the multifaceted nature of literary recognition.
Key Results
- Text-Extrinsic Dominance (Full Corpus): When analyzing the entire corpus, text-extrinsic features (especially publisher) outperformed text-intrinsic features (embeddings) in predicting canonical status. The highest F1-score for the "Canon" class was achieved using only the publisher as a feature (0.753).
- Text-Intrinsic Resilience: Embeddings alone were strong predictors, achieving an F1-score of 0.728 for the Canon class. This suggests that canonical novels do possess a distinct textual fingerprint.
- Combined Power: Combinations of text-extrinsic and text-intrinsic features generally performed comparably to the best individual feature sets, indicating a dynamic interplay rather than a simple additive effect.
- Publisher Concentration: A small number of large publishing houses (Gyldendal, Reitzel, Schubothe, etc.) were responsible for a disproportionately high percentage (94%) of canonical novels.
- False Positive Reduction: Using text-extrinsic features significantly reduced the number of non-canonical novels incorrectly classified as canonical when compared to using embeddings alone.
Limitations and Future Work
- Canon Definition: The authors acknowledge that the concept of canonicity is inherently vague and subjective. Their binary classification (Canon/Other) is a simplification, and future work could explore more continuous or multi-dimensional metrics of recognition.
- Prompt Sensitivity: The choice of prompt for the
m-e5-large-instructmodel could influence the resulting embeddings. - Training Data Bias: Web-trained models might be influenced by the online presence of canonical works, though this effect is considered minor for historical Danish literature.
- Future Directions:
- Refining the definition of canonicity with alternative lists and continuous metrics.
- Expanding the range of text-extrinsic features (e.g., critical reception, literary trends).
- Conducting more in-depth analyses of false positives and false negatives to understand the nuances of canon formation.
- Investigating the evolution of publishers' roles and their link to canonization.
- Exploring alternative embedding aggregation strategies and comparing them with simpler methods like TF-IDF.
Summary updated