Analyzing Film Adaptation through Narrative Alignment

Tanzir Pial, Shahreen Salim Aunti, Charuta Pethe, Allen Kim, Steven SkienaEMNLP 2023 · DOI

Abstract

Adapting a novel into a feature film is a complex process of selection and condensation. This paper introduces an automated framework for quantifying film adaptation by aligning book paragraphs with movie scripts. Using the Smith-Waterman local alignment algorithm combined with SBERT embeddings, the researchers identify which narrative units are retained, modified, or excised. The study provides quantitative evidence for "dialogue bias," demonstrates that narrative order is largely preserved, and introduces a metric for gender representation in adaptations.

Introduction

The transition from a 500-page novel to a 120-page screenplay necessitates significant narrative surgery. Screenwriters must decide which subplots to cut, which characters to merge, and which scenes to prioritize. Historically, analyzing these choices was a manual task for film scholars. Previous computational attempts often relied on global alignment, which fails when adaptations are "loose" or only cover portions of a source text. This research addresses these challenges by developing a robust local alignment framework to map the journey from page to screen.

Methodology: Local Alignment and Semantic Embeddings

The core of the framework is the Smith-Waterman (SW) algorithm, a technique borrowed from bioinformatics. Unlike global alignment, SW identifies the best-matching local segments between two sequences, allowing the system to ignore large deletions in the source text while finding high-confidence matches for retained scenes.

To measure similarity between book units and script scenes, the authors employ SBERT (Sentence-BERT) embeddings. This approach captures semantic meaning rather than simple keyword matching. Books are segmented into "book units" using a weighted overlap cut technique to ensure coherent context.

Visualization of narrative alignment between book units (top) and script scenes (bottom).

Datasets and Evaluation

The framework was evaluated using three distinct datasets: a manually annotated set of 6 book-script pairs for retention labeling, a manual alignment set for two Harry Potter films, and a larger automated analysis set of 40 pairs.

The SW+SBERT approach achieved an F1 score of 0.85, significantly outperforming previous methods. Notably, in split narratives like Harry Potter and the Deathly Hallows Part 2, the local alignment method correctly mapped 81.6% of the movie to the latter portion of the book, whereas global methods dropped to 41.4% accuracy because they attempted to force a match from the beginning of the text.

Key Insights: Retention and Representation

The analysis revealed several systematic trends in the adaptation process:

  • Dialogue Bias: Screenwriters are far more likely to retain book segments containing dialogue, while descriptive or internal prose is frequently excised.
  • Structural Fidelity: Despite the potential for reordering, movies almost always follow the chronological "spine" of the book. The researchers confirmed this by calculating the Longest Increasing Subsequence (LIS) of the alignments.
  • Quantifying Retention: The "retention percentage" serves as a proxy for faithfulness. As shown in the data, there is a clear negative correlation between the length of the source book and the percentage of content that makes it into the final script.
Correlation between book length and the percentage of content retained in the film adaptation.
  • The Bechdel Ratio: By analyzing the retention of dialogue between female characters, the researchers developed a "Bechdel representation ratio." This automated metric correctly predicted the official Bechdel test results for 17 out of 26 analyzed films.

Conclusion

This automated framework provides a powerful tool for the quantitative study of film adaptation. While limited by the subjectivity of alignment and copyright constraints on datasets, it opens new avenues for research. Future work may incorporate multimodal analysis, combining text with video and audio to better understand how visual elements substitute for descriptive prose.

Summary updated

BibTeX
@inproceedings{pial2023analyzing,
  author = {Tanzir Pial and Shahreen Salim Aunti and Charuta Pethe and Allen Kim and Steven Skiena},
  title = {{Analyzing Film Adaptation through Narrative Alignment}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2023},
  doi = {10.18653/v1/2023.emnlp-main.962},
}
Figure