Hate Speech Detection during Ecuador's 2022 National Strike

Published NLP Research · Springer
An applied Machine Learning study for identifying hate speech targeting Indigenous people and communities on Twitter during Ecuador's June 2022 national strike.

At a glance

Dataset	Task	Model	Notebook result
33,512 tweets	Binary hate-speech classification	TF-IDF + Multinomial Naive Bayes	84.79% accuracy

This repository is the portfolio implementation associated with the conference paper “Hate Speech Detection on Twitter: A Machine Learning Approach to Identify Attacks on Indigenous People During the 2022 Ecuador Strike.”

Why this project matters: it connects an end-to-end NLP workflow with a real research question and a peer-reviewed publication: annotation, exploratory analysis, class balancing, text processing, feature extraction, probabilistic classification, evaluation, and scientific communication.

Published research

Authors: Saire Conejo, Jairo Quelal, Silvana Escobar, Alexandra Jima-González, Erick Cuenca, José Ángel Alcántara
Publisher: Springer
Series: Lecture Notes in Networks and Systems, Vol. 1134
Pages: 267–275
DOI: 10.1007/978-3-031-70760-5_25
Publication: Springer chapter

Research question

Can Natural Language Processing and Machine Learning identify hate speech directed at Indigenous people during a major period of political and social mobilization in Ecuador?

The published study analyzes Spanish-language Twitter activity from the June 2022 national strike. The original annotated dataset contained:

• 33,512 total tweets
• 29,782 No Hate
• 3,730 Hate

Because the original classes were strongly imbalanced, the primary modeling experiment randomly undersampled the majority class to create a balanced training dataset.

ML pipeline

flowchart LR
    A["Twitter data<br/>June 2022"] --> B["Human annotation<br/>Hate / No Hate"]
    B --> C["EDA & class<br/>analysis"]
    C --> D["Class balancing"]
    D --> E["Spanish text<br/>preprocessing"]
    E --> F["TF-IDF<br/>vectorization"]
    F --> G["Multinomial<br/>Naive Bayes"]
    G --> H["Precision · Recall<br/>F1 · Accuracy"]

Loading

1. Exploratory analysis

The analysis examines class distribution and the evolution of Hate/No-Hate tweets over the strike period. The research also uses word-frequency visualization to explore recurrent language in the hate-speech subset.

2. Class balancing

The original dataset is imbalanced. The supplied notebook selects the same number of No-Hate examples as Hate examples, creating a balanced dataset before the train/test split.

3. Text representation

The classifier uses TF-IDF to transform tweet text into numerical features.

The preprocessing experiment additionally performs operations including lowercasing, removal of mentions and URLs, hashtag-symbol handling, removal of special characters, tokenization, and Spanish stop-word removal.

4. Classification

The implementation uses Multinomial Naive Bayes, a probabilistic classifier commonly used for text-classification tasks.

5. Evaluation

The supplied notebook uses an 80/20 train/test split and evaluates predictions with accuracy, precision, recall, F1-score, and a confusion matrix.

Results

Result reproduced in the supplied notebook

The baseline TF-IDF + Multinomial Naive Bayes experiment reports:

Class	Precision	Recall	F1-score
No Hate	0.88	0.79	0.83
Hate	0.82	0.90	0.86
Weighted average	0.85	0.85	0.85

Test accuracy: 84.79%

A second notebook run using the preprocessed-text column reports 84.32% accuracy, with weighted precision/recall/F1 of approximately 0.85 / 0.84 / 0.84.

Published Perspective API experiment

The paper also describes an additional verification experiment using Google Perspective API's Identity Attack attribute with a threshold of 0.5. After intersecting the human labels with this verification, the hate subset decreased from 3,730 to 213 tweets. The paper reports approximately 89.5% accuracy on the resulting 426-tweet balanced experiment.

The Perspective API implementation is not present in the notebook supplied with this repository. The 89.5% figure is therefore documented as a published-paper result, not as a result reproduced by the included code.

What the results show

The baseline experiment reaches roughly 85% overall classification performance, with particularly high recall for the Hate class in the saved notebook output. This means the model identifies a large proportion of the Hate examples in the held-out test set, while still producing false positives and false negatives.

The paper also identifies an important limitation: the approach can struggle with expressions that carry double meanings. This is especially relevant in hate-speech detection, where context, irony, coded language, and cultural nuance can affect interpretation.

Research contribution

This work demonstrates how an applied NLP pipeline can be used to investigate harmful online discourse around a real social event. It combines:

Data Collection → Human Annotation → EDA → NLP → Machine Learning → Evaluation → Scientific Publication

Personal contribution

To be completed before publishing the portfolio: add the specific parts of the research personally completed by the repository owner (for example: data collection, annotation, EDA, preprocessing, model implementation, evaluation, visualization, or manuscript preparation).

This section is intentionally not inferred from co-authorship alone.

Tech stack

Python · Pandas · scikit-learn · NLTK · TF-IDF · Multinomial Naive Bayes · Matplotlib · Seaborn · WordCloud

Repository structure

hate-speech-detection-ecuador/
├── data/
│   └── README.md
├── docs/
│   ├── hero.svg
│   └── README.md
├── notebooks/
│   └── hate_speech_detection.ipynb
├── src/
│   └── README.md
├── .gitignore
├── README.md
└── requirements.txt

Data, privacy, and ethics

The public portfolio intentionally does not redistribute the original tweet-level dataset. The research concerns abusive language targeting a marginalized community, and raw posts can contain usernames, identifiers, or harmful content that is unnecessary for demonstrating the technical workflow.

To reproduce the notebook, place an appropriately authorized dataset at:

data/hate_speech_tweets.csv

Do not commit raw tweet data, credentials, API tokens, or environment secrets.

Run locally

git clone https://github.com/YOUR-USERNAME/hate-speech-detection-ecuador.git
cd hate-speech-detection-ecuador

python -m venv .venv
source .venv/bin/activate   # Windows: .venv\Scripts\activate

pip install -r requirements.txt
jupyter notebook notebooks/hate_speech_detection.ipynb

Citation

If this research is useful to your work, please cite the published chapter:

@inproceedings{conejo2024hatespeech,
  title={Hate Speech Detection on Twitter: A Machine Learning Approach to Identify Attacks on Indigenous People During the 2022 Ecuador Strike},
  author={Conejo, Saire and Quelal, Jairo and Escobar, Silvana and Jima-Gonz{\'a}lez, Alexandra and Cuenca, Erick and Alc{\'a}ntara, Jos{\'e} {\'A}ngel},
  booktitle={Applied Engineering and Innovative Technologies},
  series={Lecture Notes in Networks and Systems},
  volume={1134},
  pages={267--275},
  year={2024},
  publisher={Springer},
  doi={10.1007/978-3-031-70760-5_25}
}

---

Research → Reproducible Code → Portfolio