Performing Gender on the Early Modern Stage

This repository contains code for data extraction, embeddings, classification, clustering and analysis for the article:

[Anonymous]. (under review). Performing Gender on the Early Modern Stage. A Computational and Cross-Lingual Approach to Male and Female Speech in European Drama.

Useful directions 📌

• european-gender-drama/ the main folder contains the Python files to extract data, get embeddings, train classifiers and run clustering experiments
• /notebooks/ contains a notebook exploring the extracted speech texts
• /data/ contains the per-language speech/gender corpora and corpus statistics
• /results/ contains results of the classification, clustering and temporal experiments
• /figs/ contains the figures generated from those results

Data & article 📝

The extracted and classified dataset will be made available upon acceptance.

Speech data is drawn from DraCor (Dutch, English, French, German, Italian and Spanish/Catalan corpora). Sentence embeddings are pushed to and loaded from HuggingFace. Links will be added later.

Project Organization 🏗️

├── README.md                        <- The top-level README for developers using this project.
│
├── requirements.txt                 <- Python dependencies.
│
├── clustering_task.py               <- Clustering task (KMeans + V-measure) to choose the best embedding model per language.
│
├── check_chunksize.py               <- Inspects how many usable speech chunks result from different chunk sizes / top-speaker cutoffs.
│
├── get_embeddings_europe.py         <- Chunk speech data and push sentence embeddings (BGE-M3 etc.) for all languages to the Hugging Face Hub.
│
├── train_classifiers_europe.py      <- Train logistic regression classifiers (TF-IDF and embeddings) to predict speaker gender, per language.
│
├── train_temporal_fre.py            <- Temporal version of the classification task: trains classifiers on 25-year subcorpora (1625–1800) of the French corpus.
│
├── add_year_to_ndjson.py            <- Adds publication year (from DraCor metadata) to the per-language NDJSON files.
│
├── add_year_to_hf_dataset.py        <- Propagates the 'year' column from the NDJSON files to the pushed Hugging Face embedding datasets.
│
├── push_to_hub.py                   <- Merges per-language local embedding datasets and pushes the combined dataset to the Hugging Face Hub.
│
├── visualize_corpus_stats.py        <- Figures on corpus composition: plays per decade, % female characters/lines, centrality by gender, speech length over time, and % female lines/characters over time for the French corpus.
│
├── visualize_method_comparison.py   <- Figures comparing classification accuracy across languages and embedding/TF-IDF methods.
│
├── visualize_temporal_results.py    <- Figures for the temporal French classification results.
│
├── data/                            <- Per-language speech/gender NDJSON files (zipped), plus corpus_stats/ (per-play and per-character CSVs) created by the notebooks.
│
├── notebooks/                       <- Jupyter notebooks.
│      │
│      └── explore_speech_texts.ipynb            <- Exploratory look at the extracted speech texts.
│
├── figs/                            <- Figures generated by the visualize_*.py scripts.
│
└── results/                         <- Results from the clustering, classification and temporal experiments (incl. v_measures/ subfolder).