Dataset

NL_restaurant_reviews

A collection of restaurant reviews was assembled in 2019 through Python-based web scraping focused on Dutch establishments, capturing both visit experiences and feature-related information. It is organized in the DatasetDict format with three splits: 116,693 training records, 14,587 test records, and 14,587 validation records.

Open on Hugging Face

Data preview

Subset

Loading preview...

Dataset card

The NL Restaurant Reviews dataset from cmotions contains roughly 146,000 Dutch-language restaurant reviews collected in 2019 via web scraping. It combines review-level text and ratings with restaurant-level identifiers and an indicator for Michelin-star status.

What this dataset contains

The corpus holds visitor-submitted reviews for Dutch restaurants, with the language tagged as Dutch and size falling in the 100K–1M row range. Each record is identified by restaurant_ID and restaurant_review_ID, and michelin_label flags establishments that are or have been Michelin-starred. Aggregate restaurant scores sit alongside individual review dimensions, making it suitable for both venue-level and review-level analysis.

Structure and fields

The dataset is distributed in Parquet under a DatasetDict layout with train, validation and test splits. Restaurant-level measures include score_total, score_food, score_service and score_decor as float64 fields. Review-level fields capture six sub-scores (reviewscore_food, reviewscore_service, reviewscore_ambiance, reviewscore_waiting, reviewscore_value, reviewscore_noise) plus the free-text review_text, review_length, and a fame_reviewer string. This pairing of structured ratings with unstructured text supports multi-modal modelling.

How analysts use it

Researchers have used the review_text feature to build classifiers predicting which Dutch restaurants are likely to earn Michelin stars, as documented in a referenced cmotions blog series. Sentiment and rating sub-scores can be aggregated to track reputation trends across venues or chains. The licensing terms are not stated, so commercial use should be confirmed with the publisher before deploying the data in production workflows.

Documentation & references 1