mananeras-word2vec

mananeras-word2vec

Spanish skip-gram Word2Vec embeddings trained on the presidential interventions of Mexican morning press conferences ("mañaneras") from December 2018 to August 2025.

Model summary

• Type: Word2Vec (skip-gram with negative sampling)
• Language: Spanish (Mexico)
• Training data: 83,138 presidential interventions from the NOSTRODATA conferencias_matutinas_amlo and conferencias_matutinas_sheinbaum repositories, preprocessed with spaCy es_core_news_md.
• Vocabulary size: 7,633 types (min_count = 10)
• Vector size: 200
• Window: 5
• Epochs: 15
• Seed: 42

Intended use

This is a research artifact for the CSCI E-89B final project “What Mexico Hears vs. What Mexico Believes” (Harvard Extension, Fall 2026). It is intended for:

• Exploring semantic neighborhoods of politically charged terms across the AMLO and Sheinbaum administrations.
• Measuring cosine displacement after orthogonal Procrustes alignment between administration-specific spaces.
• Generating t-SNE visualizations for the project report and video.

It is not a general-purpose Spanish embedding model and should not be used as a drop-in replacement for larger pre-trained embeddings.

Training pipeline

1. Read corpus_*.csv.gz from final_project/data/amlo/ and final_project/data/sheinbaum/.
2. Keep only rows where es_presidente == 1.
3. Tokenize, lemmatize, and remove stopwords, punctuation, and short tokens with spaCy es_core_news_md.
4. Train a gensim.models.Word2Vec skip-gram model.
5. Save the full gensim model (mananeras-word2vec.gensim) and the plain text vectors (mananeras-word2vec.txt).

The exact training script is in final_project/scripts/train_word2vec.py.

Limitations

• The model is trained on transcripts of public political discourse, not balanced, representative Spanish; vocabulary and associations are domain-specific.
• It reflects word usage during a particular period and administration; semantic drift is expected and is itself an object of study.
• It should not be interpreted as a truth or belief classifier.

Citation

@misc{cadena2026mananeras,
  title={What Mexico Hears vs. What Mexico Believes: Structural Topic Modeling and Transformer-Based Sentiment of Mexican Presidential Morning Conferences},
  author={Manuel Cadena Ortiz de Montellano},
  year={2026},
  note={CSCI E-89B Final Project, Harvard Extension School}
}

License and ethics

The source transcripts are public Mexican government records compiled by NOSTRODATA. This derived model is released for academic use; cite NOSTRODATA for the raw corpus.