Abstract
Movie screenplays have a distinct narrative structure that segments the story into scenes containing interleaving descriptions of actions, locations, and character dialogues. A typical screenplay spans several scenes and can include long-range dependencies between characters and events. A holistic document-level understanding of the screenplay requires several natural language processing capabilities, such as parsing, character identification, coreference resolution, action recognition, summarization, and attribute discovery. In this work, we develop scalable and robust methods to extract the structural information and character coreference clusters from full-length movie screenplays. We curate two datasets for screenplay parsing and character coreference — MovieParse and MovieCoref, respectively. We build a robust screenplay parser to handle inconsistencies in screenplay formatting and leverage the parsed output to link co-referring character mentions. Our coreference models can scale to long screenplay documents without drastically increasing their memory footprints.
Introduction: The "Final Boss" of NLP
While NLP has made massive strides in short-form text, movie screenplays present a unique challenge. They are long, semi-structured documents where a character mentioned on page 5 might not appear again until page 90. Identifying which "he," "she," or "Dr. Hager" refers to which character across an entire film is essential for automated story understanding.
Existing coreference resolution models typically struggle with two primary issues in this context:
- Structural Complexity: Screenplays consist of sluglines (scene headers), action descriptions, and dialogue. Standard parsers often break when faced with the "messy" reality of screenplay files, which frequently contain OCR errors, watermarks, or non-standard indentation.
- The Scaling Problem: Modern Transformer-based models have a limited context window. A screenplay is often 30,000 words or more. Because coreference scoring usually has quadratic complexity, standard models cannot process a whole script without running out of memory.
Methodology
The researchers proposed a two-stage solution to address these challenges.
1. Robust Screenplay Parsing
Instead of relying on fragile rule-based systems, the team built an RNN-based parser. To make it "battle-hardened," they used data augmentation, creating a "noisy" version of their dataset by synthetically adding common screenplay formatting errors like missing scene keywords, uncapitalized names, and random watermarks.
2. Scalable Coreference Resolution
They adapted a word-level coreference model and added a specific Character Scorer module to help the model distinguish between actual characters and other entities. To handle the length of full scripts, they tested two inference strategies:
- Fusion-Based Inference: The script is processed in overlapping windows. The model calculates scores for each window and averages them in the overlap zones to maintain continuity.
- Hierarchical Inference: The script is split into non-overlapping segments. The model finds clusters within each segment and then performs a second pass to merge "representative" mentions from different segments.
Datasets: MovieParse and MovieCoref
The study introduces two new benchmarks:
- MovieParse: A dataset for screenplay structural tagging, containing clean and noisy versions of 39 screenplay excerpts (over 224,000 labeled lines).
- MovieCoref: A coreference dataset consisting of six full-length screenplays (including Avengers: Endgame and John Wick) and three excerpts. This is significant because most existing coreference datasets use much shorter snippets.
Key Insights and Results
- Characters are the Anchor: Adding a dedicated character scoring module improved performance by approximately 2 points (LEA F1), suggesting that "character-ness" is a strong signal in narrative text.
- Noise-Tolerance is Mandatory: Rule-based parsers failed catastrophically on non-standard formatting (F1 scores dropping from 97% to 0.6% on certain tags), while the machine-learning approach trained on "noisy" data remained stable.
- Smart Merging Beats Brute Force: The Hierarchical and Fusion methods allowed the model to handle 30,000-word documents on standard hardware. The Hierarchical approach achieved the best overall CoNLL F1 of 71.1.
Conclusion
The authors demonstrated that increasing the "sub-document" window size and overlap significantly boosts performance by providing more context for each mention. Future work will address plural mentions and non-linear narratives (like films with heavy flashbacks) to gain deeper longitudinal insights into how movie characters interact and evolve over time.
Summary updated