AI System Turns Hotel Guest Reviews into Automated Management Decisions

A new artificial intelligence framework can read thousands of colloquial Chinese hotel reviews, work out what guests are actually complaining about, and then translate those findings directly into concrete operational decisions such as pricing adjustments, staff scheduling and facility maintenance. The system, developed by Rong Wang and described in the open-access journal Discover Artificial Intelligence, combines a Chinese-language text understanding model with a reinforcement learning algorithm, closing a loop that has long been missing in hotel analytics: the gap between knowing what guests say and acting on it automatically. The reported results are striking. Sentiment classification accuracy reached 92.37 percent, problem extraction accuracy 89.64 percent, and in a controlled comparison across three hotels of the same grade and size, the framework achieved a 96 percent problem rectification rate and a 47 percent user repurchase rate after thirty days of monitoring, well above every baseline model tested.

The problem the research addresses is familiar to anyone who has worked in hospitality. Guest reviews are the most direct signal a hotel receives about service quality, covering everything from check-in and room cleanliness to catering, facilities and staff attitude, and they shape ratings, word-of-mouth and guest retention. Previous research cited in the paper suggests that daily review sentiment has a positive impact on average daily room rate but a negative impact on occupancy rate and revenue per available room, with the trend reversing beyond a certain threshold: high sentiment scores paired with low prices are associated with increases in occupancy and revenue, while high prices only depress those metrics when sentiment is low. Yet for most mid-to-high-end hotels in China with digital systems, reviews are analysed manually, which is costly, slow and inaccurate, and the findings of that analysis rarely feed back into operational decisions in any systematic way.

Conventional keyword matching and simple sentiment dictionaries also struggle with the way guests actually write. Evaluations are fragmented into very short phrases such as remarks that soundproofing is so-so or that cleaning is perfunctory, negative sentiment is often implied rather than stated, and complaints about several service dimensions can be bundled into a single sentence, for example a leaky bathroom and an unpalatable breakfast in one composite remark. General large language models applied to the hotel industry, the paper argues, lack continuous multi-objective scheduling ability and interpretable decision links, and most existing research optimises either sentiment classification or scheduling in isolation, without a closed loop that turns text analysis into actionable operations.

The first component of the new framework, BERT-WWM, is a variant of the BERT language model pre-trained with whole-word masking, a technique particularly suited to Chinese, where text is not naturally separated into words. Instead of masking individual characters, the model masks all tokens belonging to an entire randomly chosen word, forcing it to learn complete semantic units rather than fragmented pieces. This avoids the semantic loss that character-level segmentation can cause and fits the colloquial expression style of hotel feedback, in which phrases such as high cost performance or poor hygiene appear as intact units of meaning. Positional encoding plays a supporting role here: because masking operates at the word level, the model needs a way to remember where each token sat in the original sentence. Sinusoidal position functions inject that information into the embedding vectors, with the encoding scaled to suit short texts, since hotel reviews typically run from ten to fifty words, and the embedding dimension determines how richly each token can be represented semantically.

Word order matters because in Chinese review language it can flip sentiment polarity entirely, for example when criticism arrives through contrast or a final-position negation rather than an explicit negative word. BERT-type encoders produce static, globally contextualised token vectors but do not explicitly model how meaning shifts across a sentence. The framework therefore passes the encoder output into a bidirectional long short-term memory network, or BiLSTM, which reads the review twice: once from left to right and once from right to left. The forward layer accumulates contextual information from the beginning of the sentence up to each token, while the backward layer accumulates information from the end backwards, and concatenating the two hidden states yields a fused representation capturing both global semantics and local, order-sensitive temporal dependencies. On top of this representation, the model runs two parallel business tasks: a three-class sentiment classification into positive, neutral and negative, and a five-category fine-grained problem extraction covering rooms, dining, front desk, facilities and prices.

The third component, Deep Deterministic Policy Gradient, or DDPG, is an actor-critic reinforcement learning algorithm designed for continuous action spaces. The actor network outputs a deterministic action given the current state; the critic network estimates the long-term value of state-action pairs using the Bellman equation; and experience replay plus slowly updated target networks stabilise training. In this application, the semantic feature vector of hotel reviews serves as the input state, an eight-dimensional vector whose first four dimensions hold positive and negative sentiment scores for guest rooms, catering, front desk and facilities, and whose last four dimensions hold normalised problem frequencies for each service module. The action space is four-dimensional and continuous, corresponding to pricing, manpower, maintenance and inventory preparation, each bounded within defined ranges such as a pricing multiplier between 0.8 and 1.2. A weighted reward function combines user satisfaction improvement, room revenue change and additional operating costs with weights of 0.6, 0.3 and 0.1, and a discount factor balances short-term revenue against long-term word-of-mouth and customer retention.

The resulting end-to-end framework, named BWBD, forms a continuous optimisation loop of perception, decision-making, execution and feedback. Real business labels produced by the decision module are fed back to the NLP training stage as weak supervision signals, compensating for manual labelling deviations, and the operational effects of each action generate new user feedback and new states. Ablation experiments on 600 randomly selected samples confirmed the necessity of each component: removing BERT-WWM cut sentiment classification accuracy by 9.91 percentage points and problem extraction accuracy by 10.06 points, removing BiLSTM degraded temporal modelling and hit problem extraction hardest, and removing DDPG weakened decision optimisation, with all differences statistically significant at p below 0.05. A hyperparameter sensitivity analysis found that with the discount factor fixed at 0.95, a soft update coefficient of 0.005 achieved the highest steady-state reward of 217.83 with the best balance of convergence speed and stability.

The headline numbers come from experiments on the ChnSentiCorp dataset, roughly 7,000 hotel reviews collected from publicly available user reviews on Ctrip.com and annotated manually with three levels of quality control, split 7:1:2 into training, validation and test sets and run on an RTX 4090 workstation. On 200 samples each from positive, neutral and negative subsets, the BWBD model stabilised at about 93.5 percent accuracy on positive texts, 89 percent on neutral and 92.5 percent on negative, ahead of comparison models including RoBERTa-BiLSTM-DDPG, BERT-WWM-TextCNN-DDPG, BiGRU-DDPG and LSTM-DDPG, the last of which managed only about 83 percent on positive texts. In four application scenarios, adaptation accuracy averaged 91.26 percent, peaking at 92.85 percent for front desk reception, with statistical tests against all baselines yielding p-values below 0.005.

Speed remains a trade-off. In tests with 500 sets of samples, single-text problem extraction took about 45 milliseconds for BWBD and decision response time about 49 milliseconds, a medium level that is slower than the lighter BiGRU-DDPG baseline at roughly 19 and 37 milliseconds respectively, but far faster than the RoBERTa-based comparator, which needed about 61 milliseconds for extraction and 72 milliseconds for response. Decision generation efficiency converged to around 86 percent after training, and in the 30-day controlled trial across three comparable hotels, the system’s problem rectification rate stabilised at 96 percent and its user repurchase rate at 47 percent, compared with 88 and 40 percent for the best competing model. The author is candid about the limits: the reward function covers only revenue, satisfaction and operating costs, without adaptive weighting for brand level, membership schemes or seasonality, and small changes in reward hyperparameters can destabilise pricing and staffing decisions. Training relied on a standardised simulation environment and mid-to-high-end hotel reviews, so generalisation to budget hotels or scenic-area guesthouses is untested, and offline disturbances such as sudden surges in guest flow, equipment failures or temporary staff absences are simplified.

Future work outlined in the paper includes multimodal extensions that parse images and speech to uncover service deficiencies invisible to text analysis, a multi-agent scheduling system with independent agents for rooms, staffing, catering and pricing that negotiates conflicting objectives, adaptive reward weighting by hotel type and season, and a lightweight industry large language model to make DDPG outputs traceable back to the specific negative reviews that prompted them. The paper also stresses deployment safeguards: sensitive personal data such as phone numbers and room numbers stripped during preprocessing, encrypted storage destroyed within legal time limits, feedback used only for internal service optimisation, 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. If the reported gains hold outside the simulation environment, the framework offers mid-to-high-end hotels a route from passive review reading to proactive, data-driven operations.

Subject of Research: A BERT-WWM-BiLSTM and DDPG framework that converts Chinese hotel guest feedback into automated management decisions

Article Title: Natural language processing and intelligent decision optimization technology for user feedback in hotel management

Article References: Wang, R. (2026). Natural language processing and intelligent decision optimization technology for user feedback in hotel management. Discover Artificial Intelligence, 6(1), Article 1597. https://doi.org/10.1007/s44163-026-02381-y

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02381-y

Keywords: natural language processing, reinforcement learning, BERT-WWM, BiLSTM, DDPG, hotel management, sentiment classification, user feedback, decision optimization, hospitality technology, Chinese text mining, closed-loop AI