What time is it? Temporal Analysis of Novels

Allen Kim, Charuta Pethe, Steven SkienaEMNLP 2020 · DOI

Abstract

Tracking "wall-clock time" in literature is a complex task due to the temporal sparsity of many narratives. This research presents a computational framework that combines BERT-based local classification with a global dynamic programming approach to map the temporal flow of novels. By inferring time from subtle contextual cues, the system reveals historical shifts in literary "bedtimes" and demonstrates high agreement with human annotators.

Introduction: The Challenge of Absolute Time

While human readers intuitively track the progression of time within a story, teaching machines to do so remains a significant challenge in Natural Language Processing (NLP). Most existing research focuses on relative time (e.g., "Event A happened before Event B"). This study addresses the problem of absolute wall-clock time: determining whether a scene occurs at 3:00 AM or 4:00 PM.

The primary obstacle is data scarcity. Analysis of over 50,000 books reveals that novels are often "temporally sparse." More than 10% of books contain no explicit clock-time references, and the majority contain fewer than twenty. Consequently, a robust model must move beyond simple keyword matching to infer time from contextual anchors like meals, celestial bodies, and social activities.

Methodology: From Local Cues to Global Flow

The researchers developed a multi-stage pipeline to transform raw text into a coherent temporal map:

  1. Extraction & Disambiguation: Explicit phrases (e.g., "eight o'clock") were extracted using SUTime. To resolve AM/PM ambiguity, the team trained classifiers—including Naive Bayes, LSTM, and BERT—on three-sentence windows.
  2. Local Time Prediction: The task was framed as a 24-class classification problem. This periodic approach accounts for the circular nature of time, where 11:00 PM is closer to 1:00 AM than to 1:00 PM.
  3. Global Time-Flow Analysis: To ensure narrative coherence, the authors employed Dynamic Programming (DP). This algorithm partitions the book into segments, optimizing the alignment of text to specific hours while enforcing logical temporal progression.
Temporal mapping of The Great Gatsby comparing human annotations (annot1, annot2) and the model's dynamic programming (dp) predictions.

Key Insights: Contextual Anchors and Historical Shifts

The model identifies "feature words" that serve as strong indicators of specific hours. For instance, "breakfast" peaks around 7:00–8:00 AM, while "dark" and "moon" correlate with the hours around midnight.

Top three hours associated with specific contextual anchor words, demonstrating how the model uses linguistic cues to infer time.

Beyond linguistic patterns, the framework revealed a fascinating historical trend: the "Electric Light" jump. By analyzing novels based on the author's birth year, researchers found that after 1880—the era of widespread electric lighting—the proportion of literary events occurring after 10:00 PM increased significantly. This suggests that technology fundamentally altered the "hours" kept by fictional characters.

Results and Performance

The BERT-based model demonstrated strong performance across several metrics:

  • AM/PM Disambiguation: Achieved 86.3% accuracy on the HathiTrust dataset.
  • Local Prediction: Fine-tuned BERT reached an average absolute error of 2.27 hours for specific text windows.
  • Global Mapping: The DP approach reduced the error to 4.23 hours for entire books, significantly outperforming the "Noon Baseline" (6.21 hours).
  • Human Alignment: On The Great Gatsby, the model's predictions differed from human annotators by an average of 2.62 hours—a narrow margin considering humans themselves disagreed by 1.85 hours on average.

Conclusion and Future Directions

While the current framework is highly effective, challenges remain. Optical Character Recognition (OCR) noise in older texts (e.g., "I am" being read as "1 am") can skew data. Furthermore, distinguishing between "narrator time" and "recall time" (flashbacks) requires more sophisticated narrative modeling. Future work aims to expand this "temporal fingerprinting" to include seasons, days of the week, and specific years, providing a more holistic view of time in literature.

Summary updated

BibTeX
@inproceedings{kim2020what,
  author = {Allen Kim and Charuta Pethe and Steven Skiena},
  title = {{What time is it? Temporal Analysis of Novels}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2020},
  doi = {10.18653/v1/2020.emnlp-main.730},
}
Figure