https://github.com/davidcamilo0710/hate_speech_analysis

Hate speech detection using NLP for linguistic analysis and machine learning (XGBoost) for classification with Python and SpaCy.
https://github.com/davidcamilo0710/hate_speech_analysis

hate-speech-detection linguistic-analysis nlp scikit-learn spacy xgboost

Last synced: about 1 month ago
JSON representation

Hate speech detection using NLP for linguistic analysis and machine learning (XGBoost) for classification with Python and SpaCy.

• Host: GitHub
• URL: https://github.com/davidcamilo0710/hate_speech_analysis
• Owner: davidcamilo0710
• Created: 2025-07-14T19:36:29.000Z (11 months ago)
• Default Branch: main
• Last Pushed: 2025-07-14T19:57:22.000Z (11 months ago)
• Last Synced: 2025-07-25T02:51:36.107Z (11 months ago)
• Topics: hate-speech-detection, linguistic-analysis, nlp, scikit-learn, spacy, xgboost
• Language: Jupyter Notebook
• Homepage:
• Size: 6.85 MB
• Stars: 1
• Watchers: 0
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

          
# Hate Speech Analysis and Classification with NLP and Machine Learning



This repository contains a comprehensive project for the analysis and automatic detection of hate speech in text. The project is divided into two main parts:

1.  **Linguistic Analysis (hate-analysis):** A deep exploratory analysis to identify the features and patterns that distinguish hate speech from non-hate speech.

2.  **Machine Learning Classification (hate-classification):** The development and evaluation of supervised learning models to classify messages as "Hate" or "Non-Hate" based on the extracted features.

---

## Part 1: Linguistic Analysis and Feature Extraction

In this first phase, an exhaustive analysis was conducted on a corpus of **574,639 comments** to understand the properties of hate speech. **SpaCy** and the `es_core_news_md` model were used to process the text and extract linguistic features.

### Key Findings

The comparative analysis between hate comments (2.14% of the total) and non-hate comments (97.86%) revealed significant differences:

* **Message Length:** Hate comments are drastically shorter.

    * **Average words (Hate):** 15.60

    * **Average words (Non-Hate):** 107.14

* **Message Structure:** Hate messages are more direct and contain fewer sentences.

    * **Average sentences (Hate):** 1.55

    * **Average sentences (Non-Hate):** 3.99

* **Named Entity (NER) Usage:** Hate comments tend to be less specific and more generalized.

    * **% of comments with NER (Hate):** 36.52%

    * **% of comments with NER (Non-Hate):** 59.38%

    * Specifically, only **17.59%** of hate messages mention a person (`PERSON`), compared to **28.77%** in non-hate messages.

* **Lexicon Used:** There is a clear difference in vocabulary.

    * **Hate Lemmas:** Insults ("mierda", "puta", "asco"), pejorative terms ("gentuza", "miserable"), and highly charged political words ("gobierno", "fascista", "comunista") are predominant.

    * **Non-Hate Lemmas:** The focus is on informative and neutral topics ("año", "persona", "caso", "vacuna", "salud").

* **Morphology:** It was observed that hate speech contains a higher proportion of nouns and adjectives in the **masculine plural** (18.42%) compared to the feminine plural (9.26%), suggesting a focus on male collectives.

**Conclusion of Part 1:** The analysis demonstrated that there are quantifiable linguistic features (length, structure, lexicon, etc.) that act as strong indicators for differentiating hate speech, justifying their use in building a classification model.

---

## Part 2: Hate Speech Classification

Using the findings from the first part, a Machine Learning pipeline was built to classify messages. For this task, a balanced dataset of **10,000 comments** (50% Hate, 50% Non-Hate) with pre-extracted numerical features was used.

### Models and Evaluation

Three supervised classification algorithms were trained and compared:

1.  **Random Forest Classifier**

2.  **Support Vector Machine (SVM)**

3.  **XGBoost (Extreme Gradient Boosting)**

The models were evaluated using key metrics such as **F1-Score**, **Precision**, **Recall**, and **AUC-ROC**, as it is crucial in hate speech detection to balance the correct identification of toxic messages (Recall) and the avoidance of false accusations (Precision).

### Winning Model: XGBoost

The model with the best overall performance was **XGBoost**, trained on the original (non-standardized) features.

* **F1-Score:** **0.9811**

* **Accuracy:** 98.10%

* **Precision:** 97.72%

* **Recall:** 98.50%

* **AUC-ROC:** 0.9971

This means the model is capable of **detecting 98.5% of all hate speech messages**, with **97.7% of its alerts being correct**. In practice, for every 1,000 hate speech messages, the model would only fail to identify 15.

## Results Visualization

This is an excellent place to include the graphs from the classification notebook, as they visually summarize the models' performance.

#### Performance Comparison (F1-Score)

This graph shows that **XGBoost** and **Random Forest** achieve the best performance, and that data scaling (Standardized) negatively affects SVM.



  (Insert the F1-Score comparison image here)


  



#### ROC Curves

The ROC curves demonstrate the excellent ability of all models to distinguish between the two classes, with AUC values very close to 1.

---

## Code and Technologies

* **Linguistic Analysis Notebook:** **[hate-analysis.ipynb](https://github.com/davidcamilo0710/hate_speech_analysis/blob/main/hate_analysis.ipynb)**

* **Classification Notebook:** **[hate-classification.ipynb](https://github.com/davidcamilo0710/hate_speech_analysis/blob/main/hate_clasification.ipynb)**

* **Language:** Python

* **Core Libraries:**

    * **NLP Analysis:** SpaCy, Pandas

    * **Machine Learning:** Scikit-learn, XGBoost

    * **Visualization:** Matplotlib, Seaborn