Master's Thesis · Hamburg University of Technology (TUHH) · February 2026 · Study Programme: Mechanical Engineering and Management Supervisors: Prof. Dr. Timo Heinrich · Prof. Dr. Christian Lüthje · Martin Sterner
Star ratings compress rich customer experiences into a single number. This thesis asks: what is actually being said?
Using 212,119 Google Maps restaurant reviews from Alaska, this project builds a dual-track NLP pipeline combining transformer-based sentiment classification with embedding-based topic modelling. The result: 140 semantically coherent discussion themes, each profiled by sentiment, review volume, and star-rating alignment.
Core finding: Dissatisfied customers explain; satisfied customers affirm. Negative topics are tighter, more issue-specific, and richer in detail. Positive topics cluster around generalised praise — "amazing," "delicious," "awesome."
Sentiment analysis runs first — before topic modelling — so polarity labels are never influenced by topic-specific vocabulary. Topic modelling then groups reviews by semantic similarity using dense vector embeddings (not word counts), making it robust to short, informal review language.
Note: BERTopic uses stochastic dimensionality reduction (UMAP) and density-based clustering (HDBSCAN), so topic assignments, topic counts, and sentiment distributions vary between runs — even with the same data and parameters. The results shown here are from one specific pipeline execution. The included CSV files correspond to this run.
Health and safety issues generate the most concentrated negative sentiment (93.4%), followed by interpersonal service conflicts (91.2%). Operational failures — wrong orders, long waits, confusing hours — form a consistent second tier. Negative topics are thematically sharp: customers who are unhappy explain exactly what went wrong.
Seven of the top ten positive topics hit 100% positive sentiment — but unlike negative reviews, satisfied customers rely on short affective language ("awesome," "yummy," "loved it") without detailed explanation. This is the positive–negative asymmetry: dissatisfaction motivates explanation, satisfaction motivates affirmation.
Broadly yes, but not perfectly. The most striking divergence is at 3 stars, which splits nearly 50/50 between positive and negative text. Star ratings and textual sentiment are related but not interchangeable — text adds interpretive depth that numbers alone cannot capture.
| # | Question |
|---|---|
| RQ1 | Which latent topics emerge from restaurant review text using BERTopic? |
| RQ2 | How is sentiment distributed across topics, and which themes show the highest positive or negative polarity? |
| RQ3 | To what extent does textual sentiment align with numerical star ratings? |
| Hypothesis | Result |
|---|---|
| H1: Reviews cluster into distinct recurring topics | ✅ 140 coherent topics identified |
| H2: Negative topics are more specific and explanatory than positive ones | ✅ Confirmed |
| H3: Textual sentiment aligns with star ratings but not perfectly | ✅ Strong correlation, notable mismatch at 3 stars |
| Component | Tool |
|---|---|
| Sentiment classification | distilbert-base-uncased-finetuned-sst-2-english |
| Sentence embeddings | all-MiniLM-L6-v2 (SentenceTransformers) |
| Dimensionality reduction | UMAP |
| Clustering | HDBSCAN |
| Topic modelling | BERTopic |
| Data processing | Pandas, scikit-learn |
| Visualization | Matplotlib, Seaborn |
| Hardware acceleration | CUDA · Apple MPS · CPU (auto-detected) |
.
├── README.md
├── final_code_version.py # Full analysis pipeline
├── generate_visuals.py # Chart generation script
├── requirements.txt # Python dependencies
├── data/
│ └── README.md # Dataset download instructions
├── results/
│ ├── all_topics.csv # All 140 topics with sentiment stats
│ ├── sentiment_rating_correlation.csv # Sentiment breakdown per star rating
│ ├── top_20_largest_topics.csv # 20 largest topics with examples
│ ├── top_10_positive_topics.csv # Most positive topics with examples
│ └── top_10_negative_topics.csv # Most negative topics with examples
└── visuals/
├── pipeline.svg
├── negative_topics.png
├── positive_topics.png
└── sentiment_rating_alignment.png
git clone https://github.com/Swarit786/restaurant-review-nlp.git
cd restaurant-review-nlppip install -r requirements.txtSee data/README.md for full instructions.
In short — download these two files from the McAuley Lab and place them in the data/ folder:
meta-Alaska.json— business metadatareview-Alaska.json— customer reviews
Both .json and .json.gz formats are supported.
Open final_code_version.py and set:
META_PATH = "data/meta-Alaska.json"
REVIEWS_PATH = "data/review-Alaska.json"
OUTPUT_DIR = "results/"python final_code_version.pyThe pipeline auto-detects the fastest available device (CUDA → MPS → CPU).
| Hardware | Runtime |
|---|---|
| NVIDIA GPU (RTX 3060+) | 25–45 min |
| Apple Silicon (M2/M4) | 35–60 min |
| CPU only | 2–4 hours |
python generate_visuals.pyWhy sentiment before topic modelling? Sentiment is an intrinsic property of each review, independent of thematic structure. Running it first ensures labels are unbiased by topic-specific vocabulary, enabling clean aggregation later.
Why BERTopic over LDA? LDA relies on word co-occurrence, which breaks down for short informal text. BERTopic uses contextual sentence embeddings — capturing semantic similarity even when vocabulary is sparse or varied.
Multilingual handling: Google Maps stores translated reviews as (Translated by Google) <English> (Original) <source>. The pipeline extracts only the English portion automatically.
Redundant word refinement: Inflected forms (order / orders / ordering) occupied multiple keyword slots without adding value. A targeted exclusion list was built through systematic inspection — affecting keyword representation only, not embeddings or clustering.
Institution: Hamburg University of Technology (TUHH) · Institute W-5, Digital Economics Programme: M.Sc. Mechanical Engineering and Management Submitted: February 2026 · Defended: March 2026
Yan, S., et al. (2023). Personalized Showcases: Generating Multi-Modal Explanations for Recommendations. Proceedings of the 46th International ACM SIGIR Conference. ACM. https://mcauleylab.ucsd.edu/public_datasets/gdrive/googlelocal/
Swarit Tiwari · M.Sc. Mechanical Engineering & Management · TU Hamburg LinkedIn · erswarittiwari@gmail.com