Weaponization of Migration Flows (WOM) Classification

Ever wondered how transit states turn migrant flows into a diplomatic bargaining chip? In this project, we develop and compare two NLP pipelines to automatically detect when political leaders threaten to “open the borders” or criticize destination countries’ migration policies—a tactic we term the Weaponization of Migration Flows (WOM).

Project Overview

I focus on all English‐language speeches by Turkish President Erdoğan (2014–2025), web‐scraped from the official presidency website. After manually annotating a balanced sample, I train:

• Logistic Regression + TF–IDF (with spaCy/NLTK preprocessing & SMOTE)
• Fine-tuned DistilBERT (with Hugging Face Transformers)

and evaluate both on a held-out test set as well as via 5-fold cross-validation.

Goals

• Quantify WOM by detecting “threat” or “criticism” rhetoric in presidential speeches
• Compare classical vs. transformer-based models
• Demonstrate how contextual embeddings improve detection recall (critical for catching true threats)
• Provide a reusable pipeline for analyzing WOM in other transit-migration contexts

Methods

Data Acquisition

1. Web-scraped 242 speeches → 3,976 paragraphs
2. Filtered out “visit”/“phone call” notices & salutations

Annotation

1. Hand-label 472 random paragraphs → only 16 positive WOM
2. Zero-shot classification (facebook/bart-large-mnli) → hand-verify → 338 WOM-positive + 500 WOM-negative

Baseline Model

• Preprocessing: spaCy tokenization & lemmatization + NLTK stop-word removal + custom NER tokens
• Features: TfidfVectorizer (grid-searched n-grams & min_df)
• Balancing: SMOTE
• Classifier: LogisticRegression

Advanced Model

• Preprocessing: spaCy/NLTK pipeline → BERT tokenization
• Model: Fine-tune distilbert-base-uncased for 3 epochs
• Library: Hugging Face Transformers

Evaluation

• Hold-out (20% data): Accuracy, Precision, Recall, F₁
• 5-Fold CV: mean ± std accuracy

Key Results

Model	Accuracy	Precision	Recall	F₁
Logistic Regression with TF–IDF	0.7143	0.7143	0.7224	0.7117
Advanced DistilBERT Model	0.8095	0.7195	0.8676	0.7867

• Recall Boost: +14.5 pts (72.2 → 86.8) reduces missed threats
• F₁ Improvement: +7.5 pts (71.2 → 78.7) for balanced detection
• 5-Fold CV (DistilBERT): 84.3 % ± 2.9 % accuracy (stable performance)

Repository Structure

sp2025-migration/
├── notebooks/
│   ├── wom_data_webscraping.ipynb
│   ├── wom_zeroshot_classification.ipynb
│   └── wom_nlp_model.ipynb
├── models/
│   └── wom_nlp_model.ipynb
├── README.Rmd            
└── report/
    └── report.pdf        

Further Reading

For a detailed walkthrough, see the full project report:

Alan, B. (2025). An NLP Approach to Weaponization of Migration Flows. Tulane University. [report/Report.pdf]

---

Developed by Baris Alan, Tulane University (balan@tulane.edu)