README.md

metadata

language: en
license: cc-by-nc-4.0
tags:
  - movies
  - screenplays
  - oscar
  - text-classification
  - embeddings
size_categories:
  - 1K<n<10K
task_categories:
  - text-classification
pretty_name: Movie-O-Label
dataset_info:
  features:
    - name: movie_name
      dtype: string
    - name: imdb_id
      dtype: string
    - name: title
      dtype: string
    - name: year
      dtype: int64
    - name: summary
      dtype: string
    - name: script
      dtype: string
    - name: script_plain
      dtype: string
    - name: script_clean
      dtype: string
    - name: nominated
      dtype: int64
    - name: winner
      dtype: int64

Movie-O-Label

Movie-O-Label is a dataset created by merging the MovieSum screenplay collection with Oscar nomination labels derived from David V. Lu’s Oscar Data.
It provides screenplays, summaries, and metadata together with binary labels indicating whether a movie’s screenplay received an Oscar nomination and whether it won.

---

Contents

Each entry includes:

column	type	description
movie_name	string	Title and year combined, e.g. The Social Network_2010
title	string	Movie title
year	int	Release year
imdb_id	string	IMDb identifier (e.g. tt1285016)
summary	string	Plot summary of the movie
script_clean	string	script_plain cleaned (unicode normaliziation, stage directions and scene transitions stripped where possible, whitespace reduced)
script_plain	string	Original screenplay text (only xml-tags removed from script)
script	string	Raw script field from MovieSum (for reference)
nominated	int	1 if the screenplay was nominated for an Academy Award (Writing)
winner	int	1 if the screenplay won an Academy Award (Writing)

---

Splits

The dataset is provided as a DatasetDict with:

• train — 60% (1320 movies)
• validation — 20% (440 movies)
• test — 20% (440 movies)

Splits were created stratified by the nominated label to preserve class balance.

A file split_60_20_20.npz with the exact index arrays (idx_train, idx_val, idx_test) is also provided for full reproducibility.

---

Additional Resources

To fully reproduce the experiments described in the paper:

• Paper (PDF)
• Fixed Train/Validation/Test split (split_60_20_20.npz)
• Script embeddings (emb_script.joblib)
• Summary embeddings (emb_summary.joblib)
• Title embeddings (emb_title.joblib)
• Model configuration (model_config.json)
• Projekt code jupyter notebook

License & Attribution

• MovieSum dataset:
  Created and published by Rohit Saxena (with Frank Keller).
  Licensed under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license.
  If you use this dataset, please cite:
  Rohit Saxena and Frank Keller. "MovieSum: An Abstractive Summarization Dataset for Movie Screenplays." Findings of ACL 2024. arXiv:2408.06281.

• Oscar nominations:
  Data adapted from David V. Lu!!’s Oscar Data
  Licensed under the BSD 2-Clause License © 2022 David V. Lu!!.

• Movie-O-Label:
  Created and processed by Francis Gross, based on cleaned MovieSum screenplay texts enriched with Oscar nomination and winner labels.
  Released under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license. If you use this dataset, please cite:
  Francis Gross. "Movie-O-Label: Predicting Oscar-Nominated Screenplays with Sentence Embeddings." Findings of ACL 2025 on Hugging Face.

Baseline Workflow

This work provides a simple baseline for predicting whether a screenplay receives an Oscar nomination in the Writing/Screenplays category.

1. Load the dataset

   from datasets import load_dataset
   ds = load_dataset("Francis2003/Movie-O-Label")
   

The dataset includes a predefined **60/20/20 train/validation/test split** (`split_60_20_20.npz`).

2. **Text preparation**
   Use one or more of the available feature fields:

   * `script_clean` (recommended for embeddings)
   * `summary`
   * `title`

3. **Embeddings**
   Encode the texts with [**intfloat/e5-base-v2**](https://huggingface.co/intfloat/e5-base-v2).
   Each screenplay can be chunked (e.g., 400 words with 80-word overlap), encoded,
   and mean+max pooling and L2 normalized.

4. **Classifier**
   Train a logistic regression classifier with:

   ```python
   from sklearn.linear_model import LogisticRegression
   clf = LogisticRegression(max_iter=5000, class_weight="balanced", C=1.0)
   ```

   Select the threshold on the **validation set** to maximize F1 for the positive class (nominated).

5. **Evaluation**
   Report metrics such as Accuracy, ROC-AUC, PR-AUC, F1 (positive/negative) and Macro-F1.

> The best-performing baseline used **script_clean + summary + title** embeddings
> and achieved **ROC-AUC ≈ 0.79** and **Macro-F1 ≈ 0.68** on the test set.

```


## Usage

```python
from datasets import load_dataset

# Public dataset:
ds = load_dataset("Francis2003/Movie-O-Label")

print(ds)
print(ds["train"][0])