THEaiTRE: Artificial Intelligence to Write a Theatre Play

• THEaiTRE aims to produce and stage the first computer-generated theatre play.Source: §1 · p. 1
• The project combines deep neural models trained on theatre-play data with collaboration between natural-language-processing and theatre experts.Source: §1 · p. 1
• The paper reviews previous work and data resources before outlining planned approaches and the project timeline.Source: §1 · p. 1

2.1 Narrative Natural Language Generation

• Narrative-generation research uses semantic representations, entity tracking, hierarchical planning, architectural changes, and pretrained language models to improve generated stories.Source: §2.1 · p. 1
• Most narrative-generation experiments use short, simple stories, while state-of-the-art systems still produce repetitive outputs that do not match human performance.Source: §2.1 · p. 2

2.2 Dramatic Analysis

• Theatre-study results are not generally available in formats that can be readily used as machine-learning data or annotations.Source: §2.2 · p. 2
• Dramatic-analysis research includes taxonomies of situations and character types, ontology-based annotations, and networks representing relations between characters.Source: §2.2 · p. 2

2.3 Computer-Generated Art

• Existing computer-generated artworks demonstrate technical feasibility across film, musical theatre, poetry, theatre, and music, often with substantial human involvement.Source: §2.3 · p. 2
• Mixed reception of computer-generated art indicates that current technologies are not yet on par with humans.Source: §2.3 · p. 2

• The project collects theatre scripts and synopses for training and hierarchical play generation.Source: §3 · p. 2
• More than 600 Czech plays and 1,500 synopses, alongside about 150 English plays and 100 synopses, have been collected.Source: §3 · p. 2
• The data require normalization because formats make character names, dialogue, scenic notes, and settings difficult to distinguish automatically.Source: §3 · p. 2
• The project plans to convert plays into a common JSON structure organized into acts, scenes, and dialogues, where copyright permits release.Source: §3 · p. 2

4 Planned Approach

• The project combines generative neural models, hierarchical generation, synthetic summarization, bilingual machine translation, and human-in-the-loop collaboration.Source: §4 · p. 2

4.1 Applying Neural Language Models

• GPT-2-style generation may produce theatre-like form but lacks plot progression, coherent conclusions, stable character introductions, and repetition control.The project plans fine-tuning on theatre plays and restricting which predetermined characters speak.Source: §4.1 · pp. 2–3
• Character-specific language models and explicit character embeddings, entity tracking, and coreference are planned to improve consistency and manage dialogue participants.Automatic coreference tools may require adaptation because they are typically not trained on dialogic texts.Source: §4.1 · p. 3, §4.2 · p. 3

4.2 Hierarchical Generation

• Hierarchical generation will move from titles or prompts to synopses and then to play text, with generated character lists guiding final generation.The synopsis representation may use settings, detailed external synopses, or scenic remarks, depending on data availability.Source: §4.2 · p. 3

4.3 Synthesizing Data with Summarization Techniques

• Abstractive summarization will synthesize missing synopsis levels, enabling hierarchical models to train on scripts lacking higher-level summaries.The planned sequence runs from full script to detailed synopsis, short plot synopsis, and title; Transformer models will be pretrained on news summarization and fine-tuned on theatre data.Source: §4.3 · p. 3

4.4 Machine Translation

• Machine translation will expand scarce English-Czech training data and make generated plays available in both languages, using general models fine-tuned on theatre parallel data.The project plans to translate in either direction and use its in-house Czech-English model.Source: §4.4 · p. 3

4.5 Human in the Loop

• The system will use theatre professionals to define inputs or select and rerank outputs, while only the script is autogenerated and performance remains professional.The division of labour between human and artificial components is expected to develop over time.Source: §4.5 · p. 3

5 Conclusion and Future Work

• The project officially started with preliminary work, and its proposed approaches were scheduled for implementation throughout 2020.Source: §5 · p. 3
• The first automatically generated THEaiTRE play was planned for a January 2021 premiere marking the 100th anniversary of R.U.R. [ ˇCapek, 1920].Source: §5 · p. 3
• A second play generated from an improved system was planned for January 2022, while the project could be tracked through its website.Source: §5 · p. 3