STAGE Dataset

STAGE: A Benchmark for Knowledge Graph Construction, Question Answering, and In-Script Role-Playing over Movie Screenplays

A bilingual (Chinese and English) benchmark for narrative understanding over movie screenplays.

Dataset Overview

STAGE contains structured annotations for understanding complex narratives in both Chinese and English films. The dataset includes:

• Chinese: 42 Chinese movie scripts with full annotations
• English: 109 English movie scripts with full annotations
• Metadata: chinese_movie_info.csv and english_movie_info.csv with movie-level information (title, word count, number of scenes, genres)

Directory Structure

Each movie directory contains:

File / Directory	Description
script.json	Scene-by-scene segmented movie script
episodes.json	Episode-level narrative units with descriptions and related events
episde_relations.json	Pairwise temporal/causal relations between episodes with confidence scores
episde_causality_graph.json	Weighted causality graph over episodes
extraction_results.json	Per-scene entity and event extraction results
doc2chunks.json	Document-to-chunk mapping with full text and metadata
rename_map.json	Entity name normalization/alias map
question_pairs.csv	Multi-hop question-answer pairs with evidence and QA type annotations
ICRP/	In-Character Role-Playing data (per-character subdirectories)

ICRP (In-Character Role-Playing)

Each movie's ICRP/ directory contains one subfolder per character (typically 3 characters per movie), with:

File	Description
persona_card.json	Character traits, speaking style, behavioral constraints, and dialogue exemplars
key_relations.json	Key relationships with other characters
icrp_qa.json	Role-playing QA pairs with supporting/contradicting facts
actions.csv	Character's physical actions per scene
dialogues.csv	Character's dialogue lines per scene
facts.csv	Factual statements about the character per scene

Data Format

script.json

[
  {
    "_id": 1,
    "title": "Scene Title",
    "subtitle": "Scene Subtitle",
    "content": "Scene content with dialogue and actions..."
  }
]

episodes.json

[
  {
    "id": "ep_4f8f0b6d18e73a34",
    "name": "Episode Title",
    "description": "Narrative description of the episode",
    "source_documents": ["scene_7_part_1"],
    "related_events": ["event description 1"],
    "related_occasions": ["occasion description"]
  }
]

episde_relations.json

[
  {
    "id": "rel_ep_ep_8f935bc28619",
    "subject_id": "ep_...",
    "object_id": "ep_...",
    "relation_type": "precedes",
    "reason": "Natural language justification",
    "source_documents": ["scene_11_part_1"],
    "confidence": 0.95
  }
]

episde_causality_graph.json

[
  {
    "subject_id": "ep_...",
    "object_id": "ep_...",
    "relation_type": "CAUSAL_LINK",
    "original_relation_type": "precedes",
    "weight": 0.7,
    "effective_weight": 0.595
  }
]

question_pairs.csv

Column	Description
id	Question ID
scene	Related scene
question	Question text
answer	Answer text
evidence	Supporting evidence from the script
qa_type	QA type (e.g., character states, causal/relational queries, dialogue/beliefs, detailed description, temporal references)

Publication Status

Published as arXiv preprint

Citation

If you use this dataset in your research, please cite our paper:

@article{tian2026stage,
  title={STAGE: A Benchmark for Knowledge Graph Construction, Question Answering, and In-Script Role-Playing over Movie Screenplays},
  author={Tian, Qiuyu and Li, Yiding and Chen, Fengyi and Liu, Zequn and Kong, Youyong and Guo, Fan and Li, Yuyao and Shen, Jinjing and Xie, Zhijing and Luo, Yiyun and others},
  journal={arXiv preprint arXiv:2601.08510},
  year={2026}
}

Contact

For questions or issues regarding the dataset, please open an issue in this repository.