IndoHotelABSA: A Large-Scale Aspect-Based Sentiment Analysis Benchmark for Indonesian Hotel Reviews

Published: 18 August 2026| Version 2 | DOI: 10.17632/9vhzg5wkf9.2
Contributors:
,
,
,
,

Description

IndoHotelABSA is a large-scale benchmark for aspect-based sentiment analysis (ABSA) of Indonesian hotel reviews. It contains 14,988 unique, deduplicated Indonesian reviews collected from the Google Places API across 55 cities spanning all six major regions of Indonesia (Sumatra; Java; Bali & Nusa Tenggara; Kalimantan; Sulawesi; Maluku & Papua). All personally identifying metadata were removed at collection; only review text, star rating, hotel name, and city are retained. A 3,000-review, class-balanced sample was annotated for seven CRM-meaningful aspect categories (Location; Cleanliness; Service/Staff; Room & Facilities; Price/Value; Food & Beverage; Public Facilities), each with three-way polarity (positive/negative/neutral), using a two-tier, LLM-assisted, human-verified protocol. A 500-review gold subset was independently verified by three annotators at almost-perfect agreement (Fleiss' kappa = 0.894; 92.1% of LLM drafts retained, 7.9% corrected), yielding 1,506 gold aspect annotations. The release distinguishes LLM-drafted (silver) from human-verified (gold) labels and provides fixed train/validation/test splits (2,480 silver train; 100 gold validation; 400 gold test). All reported evaluations in the accompanying article use the human-verified gold labels. Version 2: licence now stated per component (annotations CC BY 4.0, review text redistributed as third-party content, not relicensed). In-text PII scrubbing applied. Per-annotator label files, adjudication record, gold subset identifiers, corpus id mapping, blinded re-annotation, and the 20 excluded records added. Documentation corrected so that "human-verified" refers only to the 500-review gold subset.

Files

Steps to reproduce

- Companion research article: "IndoHotelABSA: An Explainable Aspect-Based Sentiment Analysis Framework and Benchmark Dataset for Indonesian Hotel Reviews in Hospitality CRM," Natural Language Processing Journal (add DOI on acceptance). - Companion data descriptor: Data in Brief (add DOI when available).

Institutions

Categories

Computational Linguistics, Natural Language Processing, Sentiment Analysis, Low-Resource LLM

Funders

  • Universitas Dinamika Bangsa

Licence