AI turns Chinese hotel reviews into management decisions

In a 30-day controlled comparison across three comparable mid-to-high-end hotels in China, an experimental AI system posted a 96% problem-rectification rate and a 47% user repurchase rate, versus 88% and 40% for the best competing model.

Those figures came from a research test, not a documented chain rollout. Rong Wang's 2026 Discover Artificial Intelligence paper examines an AI framework that turns short Chinese reviews into pricing, staffing, maintenance and inventory decisions. The study names no chain, ownership group, franchise structure or commercial payer.

How the system connects complaints to operations

BWBD joins three components instead of separating review analysis from hotel scheduling. BERT-WWM interprets Chinese, BiLSTM captures sentence-order effects, and DDPG selects continuous operating actions. Wang says general LLMs lack continuous multi-objective scheduling and interpretable decision links, while earlier work optimizes sentiment or scheduling in isolation.

BERT-WWM masks whole words rather than individual characters, reducing semantic loss in a language without spaces between words. Sinusoidal position encoding retains token location, with scaling suited to the dataset's 10-to-50-word reviews.

BiLSTM reads each review forward and backward, catching polarity changes caused by contrast or final-position negation. It still handles implied complaints such as “soundproofing is so-so” and “cleaning is perfunctory,” and separates bundled issues such as a leaky bathroom and an unpalatable breakfast. Conventional keyword matching and sentiment dictionaries often miss those patterns.

Two parallel outputs classify sentiment as positive, neutral or negative and extract problems in rooms, dining, front desk, facilities and prices.

DDPG supplies the decision engine. It is an actor-critic algorithm: the actor proposes a deterministic action, while the critic estimates long-term state-action value using the Bellman equation. Experience replay and slowly updated target networks stabilize training.

• Input state: An eight-dimensional vector combines positive and negative sentiment scores for guest rooms, catering, front desk and facilities with normalized problem frequencies for each service module.
• Actions: The system can adjust pricing, manpower, maintenance and inventory preparation. A pricing multiplier ranges from 0.8 to 1.2, effectively allowing room rates to fall 20% or rise 20%.
• Reward: Satisfaction improvement carries a 0.6 weight, room-revenue change carries 0.3, and additional operating costs carry 0.1. A discount factor balances near-term revenue against long-term word-of-mouth and retention.
• Loop: Real business labels return as weak supervision, compensating for manual-labeling deviations; actions generate new feedback and states, creating perception, decision, execution and feedback.

What the tests show

The model reached 92.37% sentiment-classification accuracy and 89.64% problem-extraction accuracy. Researchers trained it with ChnSentiCorp, roughly 7,000 reviews pulled from public Ctrip.com posts and manually annotated through three levels of quality control. They split the data 7:1:2 for training, validation and testing and ran it on an RTX 4090 workstation.

On 200 samples each from positive, neutral and negative groups, BWBD reached about 93.5%, 89% and 92.5% accuracy, respectively. It beat RoBERTa-BiLSTM-DDPG, BERT-WWM-TextCNN-DDPG, BiGRU-DDPG and LSTM-DDPG; the last managed only about 83% on positive texts.

Across four application scenarios, adaptation accuracy averaged 91.26% and peaked at 92.85% for front-desk reception, with p-values below 0.005 against all baselines.

Across 500 sample sets, extraction took about 45 milliseconds and decision response about 49 milliseconds, versus roughly 19 and 37 milliseconds for lighter BiGRU-DDPG. The RoBERTa-based comparator needed about 61 and 72 milliseconds. Decision-generation efficiency converged near 86% after training.

Ablation tests on 600 randomly selected samples showed why each layer matters. Removing BERT-WWM cut sentiment accuracy by 9.91 percentage points and extraction accuracy by 10.06 points; removing BiLSTM degraded temporal modeling and hit extraction hardest, while removing DDPG weakened decision optimization. The study found all differences significant at p below 0.05.

With the discount factor fixed at 0.95, a 0.005 soft-update coefficient produced the highest steady-state reward, 217.83, with the best balance of convergence speed and stability.

What does this change for hotel operators?

It converts guest feedback from a lagging report into a control signal for prices, labor and upkeep. That shift matters because prior research cited in the paper linked daily review sentiment to higher average daily room rate but lower occupancy and revenue per available room, until a threshold reversed the pattern.

High sentiment and low prices correlated with occupancy and revenue gains, while high prices hurt those measures when sentiment was weak. The 47% repurchase rate therefore measures retention, not profit.

Yet the study reports no property-level occupancy, ADR, RevPAR, gross margin, cost of goods, food cost, labor percentage, implementation cost, licensing terms or payback. It also does not identify whether an owner, operator, management company or franchise group pays. The reward function omits brand level, membership programs and seasonality.

What remains before deployment

Hyperparameter changes can destabilize pricing and staffing, and the work relies on a standardized simulation environment. Generalization to budget hotels and scenic-area guesthouses remains untested, while guest surges, equipment failures and temporary staff absences receive simplified treatment.

Future work includes image and speech analysis for complaints hidden in text, and independent agents for rooms, staffing, catering and pricing that negotiate conflicting objectives. It also proposes adaptive reward weighting by hotel type and season, plus a lightweight industry large language model that traces DDPG decisions back to specific negative reviews.

Wang says deployment needs “hard upper and lower bounds on pricing and staffing actions” and “regular human review of algorithmic outputs to prevent revenue-driven overpricing or cuts to service manpower.” The proposed safeguards also strip phone and room numbers during preprocessing, store encrypted data for destruction within legal limits, restrict feedback to internal service optimization and require human oversight.

The next test is whether live hotels reproduce the reported gains under real demand and operating shocks without sacrificing service levels.