kwongnon/2026-24679-text-dataset

Hospitality Reviews: Hotel vs Restaurant

kwongnon/2026-24679-text-dataset

English-language hospitality reviews labeled by venue type. The classification task is to predict whether a review describes a hotel or a restaurant. Labels are derived from the Hospitality column: 0 = restaurant; 1 = hotel. The target describes the venue category, not review sentiment, review quality, or whether the statements in a review are factually correct.

Source and task

The preparation notebook reads the tabular review data, retains Review and Hospitality, creates a stable source_id, and converts the category into numeric and readable labels. Each source row is one classification example. The packaged training split also contains synthetic variants generated only from original training reviews.

No author, customer, hotel, or restaurant identity is used as a grouping key. If multiple reviews refer to the same person or establishment, the current row-level split does not guarantee that they remain in the same partition.

Fields

Splits and class counts

These counts are computed from the packaged splits for this run.

Class counts include synthetic rows in training only.

Requested holdout fraction: 30%; test receives 50% of that holdout. The original rows are split with stratification by label. The first split uses seed 24679, and the holdout split uses seed 24680.

Split original reviews before augmentation. Every generated review remains in training with its parent. Validation and test contain only original reviews. Reuse these fixed partitions for all model comparisons, and do not randomly re-split the augmented training pool.

Augmentation and preprocessing

The packaged training pool targets a 15x multiplier: one original plus 14 generated variants for every original training review. Repeated augmentation runs may reuse the same method with different deterministic seeds. Every run starts from the original review; transformations are not stacked on previously generated text.

Character edit rate: 8% of eligible words, capped at 3 edited words per generated review.

• —Internal character swaps: exchange adjacent, distinct interior letters in selected words.
• —Internal character deletions: remove one interior letter from selected words.
• —Internal character duplications: duplicate one interior letter in selected words.
• —Word case changes: alter the case of selected words without changing their letters.
• —Within-line word swaps: exchange two distinct words while preserving each line's word inventory.
• —Word duplications: repeat selected words to simulate natural typing or editing noise.
• —Punctuation removals: remove selected punctuation marks.
• —Punctuation insertions: insert punctuation after selected words.
• —Mixed typo noise: combine internal character swaps and duplications.

Only nonempty variants that differ from their parent and from earlier variants of the same parent are accepted. Labels and venue categories are copied from the original parent. The exact-multiplier generation routine retries with deterministic alternative seeds and stops with an error if it cannot produce the required number of unique changed variants; it never counts an unchanged copy as augmentation.

The coverage table reports scheduled runs, generated rows, and distinct parents covered by each method. Structural checks verify parentage, copied labels, nonempty changed text, unique IDs, and within-line word preservation for word swaps. These checks do not prove that a generated review still expresses its assigned venue category, so representative variants from both labels must also be reviewed manually.

Intended use and limitations

Use this dataset for teaching or evaluating binary text classification, data preparation, and text augmentation. Report validation and test class counts alongside accuracy, precision, recall, and F1, and compare augmented training against original-only training.

The data may contain strong lexical shortcuts such as room, front desk, menu, or server. Models can therefore learn venue-specific vocabulary rather than general hospitality concepts. Row-level random splitting may also overestimate generalization if reviews from the same establishment occur in multiple partitions. The augmentation operations simulate surface noise; they do not create new establishments, experiences, or independent observations. Some variants may be ungrammatical or may alter meaning despite retaining the parent's label. Human review remains necessary.

Do not interpret label as a sentiment score, quality rating, authorship judgment, or factual verification. Do not use this convenience dataset for consequential decisions about customers, workers, or businesses.

Privacy and licensing

Free-text reviews may contain names, locations, travel details, or other identifying information. Pseudonymous source IDs do not establish anonymity or consent. Review the source data and applicable permissions before redistributing or publishing the dataset.

No license is assigned by this dataset card. Preserve and follow the terms of the original data source and document them before public release.

Load and compare

from datasets import load_dataset

ds = load_dataset("kwongnon/2026-24679-text-dataset")
# Train with ds["train"], tune with ds["validation"], and evaluate once on ds["test"].

If repository visibility changes, use an account with access. For reproducible comparisons, record the dataset commit, split seeds, augmentation configuration, and model/environment versions. Regenerate this card after changing the source data, split, or augmentation settings so that its tables continue to describe the packaged files.