README.md

---
language:
  - en
pretty_name: "Plot Twists Over Time: How Movie Stories Have Changed Over 95 Years"
tags:
  - movies
  - embeddings
  - semantic-analysis
  - temporal-analysis
  - text-embeddings
  - plot-summaries
  - genre-classification
  - concept-extraction
  - wikidata
  - wikipedia
  - tmdb
  - bge-m3
  - text
license: "cc-by-nc-4.0"
task_categories:
  - text-classification
  - sentence-similarity
  - text-retrieval
  - feature-extraction
size_categories:
  - 10K<n<100K
dataset_info:
  features:
    - name: movie_id
      dtype: string
    - name: title
      dtype: string
    - name: year
      dtype: int32
    - name: plot
      dtype: string
    - name: genre
      dtype: string
    - name: embedding
      dtype: float32
      shape: [1024]
---

# Movie Plot Embeddings Dataset

**Project Title**: Plot Twists Over Time: How Movie Stories Have Changed Over 95 Years

## Dataset Summary

This dataset contains movie metadata, plot summaries, and semantic embeddings for ~92,000 feature-length films (1930-2024). Created for temporal semantic drift analysis, it includes metadata from Wikidata, TMDb, and Wikipedia, along with dense (1024-dim) and sparse embeddings generated using BAAI/bge-m3.

**Key Statistics**: ~92,000 movies | Year range: 1930-2024 | Embedding dimension: 1024 | Average plot length: ~1,500 characters

## Dataset Structure

### Core Files

- **`final_dataset.csv`** (200 MB): Movie metadata with 39 columns (title, year, genre, plot, directors, actors, TMDb metrics, etc.)
- **`final_dense_embeddings.npy`** (380 MB): Dense embeddings array `(N, 1024)` using BGE-M3 CLS token aggregation
- **`final_dense_movie_ids.npy`** (7 MB): Wikidata QIDs corresponding to embeddings (index-aligned: `embeddings[i]` ↔ `movie_ids[i]`)
- **`final_sparse_embeddings.npz`** (500 MB): Sparse lexical weights for concept extraction (token_indices, weights, movie_ids)

### Additional Files

- **`knn_faiss_novelty.csv`**: Novelty scores and neighbor information
- **`umap_cluster_trajectories.png`**: UMAP visualization of embeddings
- **`concept_space/`**: WordNet-based concept vocabulary and embeddings for semantic mapping

## Data Collection Pipeline

Four-step pipeline: (1) **Wikidata**: Query SPARQL for movie metadata (1930-2024, ~8K/year), filter feature films; (2) **TMDb**: Enrich with popularity, votes, ratings via API; (3) **Wikipedia**: Extract plot summaries from sitelinks; (4) **Embeddings**: Generate dense/sparse embeddings using BGE-M3 with CLS token aggregation, parallel GPU processing.

## Data Fields

**Core**: `movie_id` (Wikidata QID), `title`, `year`, `imdb_id`, `tmdb_id`

**Metadata**: `release_date`, `country`, `duration`, `wikidata_class`, `wikipedia_link`

**Creative Team**: `directors`, `directors_id`, `actors`, `actors_id` (pipe-separated for multiple)

**Genre**: `genre` (raw, comma-separated), `genre_id`, `genre_cluster_ids`, `genre_cluster_names` (processed)

**Plot**: `plot` (full text), `plot_section`, `plot_length_chars`, `plot_length_tokens`, `num_different_tokens`, `token_shannon_entropy`

**TMDb Metrics**: `popularity`, `vote_average`, `vote_count`

**Financial** (if available): `budget`, `budget_currency`, `box_office`, `box_office_currency`, `box_office_worldwide`

**Other**: `set_in_period`, `awards`, `summary`

## Loading the Dataset

### Basic Python Example

```python
import numpy as np
import pandas as pd

# Load metadata and embeddings
df = pd.read_csv('final_dataset.csv', low_memory=False)
embeddings = np.load('final_dense_embeddings.npy')
movie_ids = np.load('final_dense_movie_ids.npy', allow_pickle=True)

# Merge embeddings with metadata
embeddings_df = pd.DataFrame({'movie_id': movie_ids, 'embedding': list(embeddings)})
combined_df = pd.merge(df, embeddings_df, on='movie_id', how='inner')
```

### Using Utilities

```python
from src.data_utils import load_final_data_with_embeddings

df = load_final_data_with_embeddings(
    csv_path="data/data_final/final_dataset.csv",
    data_dir="data/data_final",
    verbose=True
)
```

## Embedding Details

- **Model**: BAAI/bge-m3 | **Dimension**: 1024 | **Method**: CLS token aggregation | **Normalization**: L2-normalized
- **Alignment**: `final_dense_embeddings[i]` corresponds to `final_dense_movie_ids[i]`
- **Alternative chunking**: Supports mean pooling, chunk-first, late chunking (stored with suffixes)

## Concept Space

Pre-computed WordNet-based concept vocabulary: top 20K nouns (Zipf ≥ 4.0), embedded with BGE-small-en-v1.5. Files parameterized by Zipf threshold, vocabulary size, and model name.

## Data Quality

**Completeness**: All movies have core fields; ~92K plots, ~91K TMDb data, ~75K directors, ~81K genres.

**Filtering**: Feature films only; plots filtered by length/entropy; explicit content excluded; duplicates removed.

**Cleaning**: Plot text normalized; genres clustered; Shannon entropy threshold: 4.8398.

## Use Cases

**Temporal Analysis**: Compute decade centroids to analyze semantic drift over time.

**Genre Classification**: Use embeddings for clustering or classification tasks.

**Similarity Search**: Find similar movies using cosine similarity on embeddings.

**Concept Extraction**: Map plot nouns to concept space using sparse lexical weights.

## Citation

```bibtex
@dataset{movie_plot_embeddings_2026,
  title={Plot Twists Over Time: How Movie Stories Have Changed Over 95 Years},
  author={Cheung, Ansel and Villa, Alessio and Markovinović, Bartol and López de Ipiña, Martín and Abraham, Niklas},
  year={2026},
  note={Dataset collected from Wikidata, TMDb, and Wikipedia for temporal semantic analysis of movie plots}
}
```

## License

This dataset inherits licenses from source data:
- **Wikidata**: CC0 1.0 (Public Domain)
- **Wikipedia**: CC BY-SA 4.0
- **TMDb**: CC BY-NC 4.0 (non-commercial)

**Dataset License**: CC BY-NC 4.0 for non-commercial research. Commercial use requires TMDb licensing and Wikipedia compliance.

**Code**: MIT License (see main repository).

**Attribution Required**: Cite Wikidata, Wikipedia, TMDb contributors, and this dataset.

## Acknowledgments

Wikidata, TMDb, Wikipedia, BAAI (BGE-M3), WordNet

## Notes

- Embeddings are L2-normalized
- Movie IDs are Wikidata QIDs (format: "Q####")
- Plot text cleaned and normalized
- Genres may be multi-label (pipe-separated)
- Some fields may be NaN for older/less popular films