This project explores 30,000+ online reviews of running shoes written in Italian, French, and English.
Using Natural Language Processing (NLP) and BERTopic, the goal is to understand what customers talk about most, from comfort and fit to quality, design, and value for money.
It’s an end-to-end multilingual pipeline covering:
Data cleaning → Semantic embeddings (MPNet) → Topic modeling → Evaluation → Human interpretation.
- Identify key themes in customer reviews across languages
- Compare linguistic and sentiment patterns
- Turn unstructured feedback into actionable business insights
BERTopic Pipeline
The figure above summarizes the full process:
- Web Scraping from public e-commerce pages (Decathlon, Amazon, Zalando)
- EDA & Cleaning to handle duplicates, encoding errors, and detect language
- Word Embeddings with MPNet
- Dimensionality Reduction using UMAP
- Clustering with HDBSCAN
- Tokenizer & Weighting via c-TF-IDF (BERTopic)
- Evaluation & Fine-Tuning for coherence and interpretability
- Results & Insights for customer understanding
| Step | Description |
|---|---|
| EDA & Cleaning | Handle missing values, encoding errors, and language detection |
| Preprocessing | Multilingual stopwords, normalization, text cleaning |
| Embeddings | SentenceTransformer (all-mpnet-base-v2) |
| Topic Modeling | UMAP + HDBSCAN + BERTopic |
| Evaluation | Coherence (c_v), silhouette, % noise |
| Topic Reduction | Merge similar topics into macro-categories |
- Size: ~30k reviews
- Languages: 🇮🇹 Italian, 🇫🇷 French, 🇬🇧 English
- Sources: Decathlon (52%), Amazon (36%), Zalando (12%)
- Timeframe: Nov 2024 – Feb 2025
- Fields: Product info, technical specs, ratings, and text reviews
| Category | Description | Share |
|---|---|---|
| Comfort & Fit | How shoes feel, adapt, and support | ~50% |
| Performance & Cushioning | Running efficiency and energy return | ~20% |
| Quality & Durability | Longevity and material quality | ~10% |
| Design & Style | Aesthetic preferences and branding | ~8% |
| Price & Value | Perceived fairness and expectations | ~7% |
| Purchase Experience | Delivery, sizing, returns | ~5% |
Model performance:
- Coherence: 0.68 (after topic reduction)
- Silhouette score: 0.38
- Noise: ~15%
- Comfort and fit are the strongest drivers of satisfaction across all languages.
Customers consistently associate comfort, foot support, and fit precision with positive ratings. - Durability and quality emerge as the most frequent negative aspects, especially for high-performance models.
- Price fairness appears in both positive and negative reviews; people praise good deals but criticize overpriced products.
- French reviews often discuss value-for-money and aesthetic appeal, while Italian reviews include more contextual and emotional tone (e.g. “great for daily runs”).
- English reviews are shorter and more direct, focusing mainly on comfort and durability.
- Highlight “comfort-driven design” in product storytelling.
- Invest in durability communication (materials, lifespan).
- Localize content:
- 🇮🇹 Focus on emotional and lifestyle tone
- 🇫🇷 Emphasize quality/price ratio
- 🇬🇧 Keep concise, technical information
- Monitor recurring mentions of sizing issues and wear resistance for product feedback loops.
Languages & Tools:
Python 3.11 · pandas · numpy · matplotlib · seaborn · spaCy · nltk · ftfy
sentence-transformers · BERTopic · umap-learn · hdbscan · gensim · scikit-learn
- Short reviews: Most comments are under 30 words, limiting topic coherence.
- Language imbalance: English reviews dominate, slightly biasing results.
- Hardware constraints: The analysis was run on CPU, preventing large-scale fine-tuning.
- Sentiment approximation: Ratings were used as sentiment labels (not direct text classification).
- Temporal gap: Reviews were scraped over different months, so no consistent time trend was analyzed.