💬 Cross-Linguistic NLP Analysis of E-Cigarette Perceptions on Social Media

Python | Hugging Face | BERTweet | Twitter Twin BERT | Sentence-BERT | UMAP | K-Means | BERTopic | Multilingual NLP | Public Health AI

---

🌍 Overview

Led a multilingual NLP capstone project analyzing 51K English and 7K Spanish tweets (sampled from 1.1M+ total) to uncover cultural and linguistic differences in public e-cigarette discourse.
The project integrated Transformer-based models and BERTopic clustering to reveal how cultural and linguistic nuances shape online conversations about vaping and health.

Goal: Bridge the gap between multilingual social media analysis and public health strategy by uncovering culturally specific narratives in e-cigarette discussions.

---

🧩 Dataset & Preprocessing

• Data Source: Third-party provider (keyword-based Twitter/X scraping)
• Dataset Size:
  • 1.1M English tweets → 51K after cleaning
  • 34K Spanish tweets → 7K after cleaning
• Preprocessing Pipeline:
  • Deduplication and retweet removal
  • Full-text crawling for truncated tweets
  • Tokenization, lemmatization, emoji/special character filtering
  • Text normalization for both English and Spanish corpora

---

🧠 Model Training & Fine-Tuning

Built and fine-tuned multiple Transformer-based classifiers to detect multilingual relevance, commercial intent, and attitude toward e-cigarettes.

• Relevance Classification:
  → Model: BERTweet
  → Accuracy: 96%, F1: 0.85

• Commercial Intent Detection:
  → Model: BERTweet
  → Accuracy: 93%, F1: 0.87

• Attitude Classification (Positive / Neutral / Negative):
  → Model: Twitter Twin BERT Large
  → Accuracy: 92%, F1: 0.89

• Manual Labeling: 1,000 tweets labeled through iterative consensus across annotators.
• Experiment Tracking: Comparative evaluations conducted via Hugging Face Transformers and custom Python scripts.

---

🔍 Topic Modeling Framework

• Pipeline: Sentence-BERT embeddings → UMAP (dim. reduction) → K-Means clustering
• Model Comparison: TF-IDF, LDA, and BERTopic evaluated
• Final Selection: BERTopic for best coherence and topic diversity
• LLM Integration: OpenAI-based LLMs used for iterative topic naming (100+ refinements)

---

📊 Key Findings

• Cross-Linguistic Insights:
  • English tweets: Focused on regulation, youth vaping, and health risks
  • Spanish tweets: Emphasized social lifestyle, recreation, and cultural expression
  • Translation revealed sentiment dilution — idioms and slang often lost nuance
• Sentiment vs. Attitude:
  • Traditional sentiment (VADER): Majority neutral
  • Fine-tuned attitude model: 41.2% negative, 24.6% positive, 34.2% neutral — exposed hidden polarization
• Translation Drift:
  • Spanish originals: addiction, cravings, habit discussions
  • Translations: introduced consumer and gifting themes not found in native tweets

---

⚙️ Challenges

• Subjective Labeling: Cross-cultural definitions of “positive/negative” required careful alignment
• Translation Fidelity: Idioms, emojis, and slang were inconsistently captured in MT pipelines
• Dataset Imbalance: English tweets outnumbered Spanish by ~7:1
• Cross-Language Topic Coherence: Syntactic variations reduced cluster interpretability
• Demographic Bias: Twitter users may not represent broader population sentiment

---

🚀 Future Work

• Develop cross-lingual topic alignment models to reduce translation drift
• Expand labeled data for better multilingual fine-tuning
• Perform temporal sentiment tracking around policy updates and major public health events

---

📄 Full Report: E-Cigarette Perception Analysis (PDF)

---

📘 Keywords:
Multilingual NLP | BERTweet | Sentence-BERT | BERTopic | Cross-Linguistic Analysis |
Transformer Models | Public Health | UMAP | Clustering | Cultural Interpretation