Qunar Intelligent Service Robot: Architecture, Cognitive System, and Iterative Development

Since the advent of computers, engineers have used languages like Python and Java to automate slow, manual processes, dramatically boosting productivity. Facing a 50% surge in service‑call volume during peak travel periods, Qunar decided to build its own AI service robot rather than rely on additional human agents.

The existing manual service data shows a sharp peak in call volume, with most inquiries related to tickets, hotels, and trains. In the hotel domain, the majority of questions concern order status, cashback rules, and invoicing.

Traditional call handling follows a four‑step workflow: answer the phone, locate the user's order in the backend system, retrieve the relevant business rules, and explain the answer to the user.

With the robot, users submit questions via the app (order details, service channel). The input is processed by the q‑robot‑brain cognitive system, which determines intent and generates a response.

The overall product structure is illustrated in a diagram (image). The system abstracts user queries, automatically retrieves user and product status information, and then delivers the appropriate answer.

The knowledge base consists of two layers: a static FAQ layer with unchanging answers, and a dynamic layer that varies with business context (e.g., refund policies differ for direct‑booked versus supplier‑booked hotels). Dynamic knowledge requires real‑time API calls and timely updates to avoid misleading users.

Two subsidiary subsystems support the robot: an effectiveness‑evaluation system that monitors service volume, success rates, and the proportion of users still opting for human agents; and a labeling system that lets operations staff review robot answers, annotate correct or incorrect responses, and provide valuable dialogue data for model training.

Since its launch in December 2016, the robot has accumulated over 30,000 chit‑chat utterances and more than 1,000 business‑specific Q&A items covering flights, hotels, trains, vacations, and tickets. It serves roughly 20,000 daily users, achieving a self‑service rate above 60% and handling about 8% of the total order‑related service volume.

The cognitive system’s performance hinges on knowledge‑base coverage and matching accuracy. It has undergone three major iterations:

Phase 1 employed Elasticsearch retrieval to match user queries to standard questions, enabling a rapid MVP within one month.

Phase 2 added NLP enhancements such as stop‑word removal, synonym expansion via word2vec, and an annotation platform to increase coverage and collect labeled data.

Phase 3 rebuilt the architecture, introducing intent recognition, a small‑talk module, and a convolutional neural‑network (CNN) for semantic matching.

In the intent‑recognition module, two classifiers run in parallel: an SVM and a rule‑based classifier that flags business keywords. Business keywords are extracted using TF‑IDF, expanded with word2vec, and stored in a Bloom filter.

The small‑talk module uses a retrieval‑based approach (pre‑defined Q&A pairs) for greetings and thank‑you messages; if no match is found, the query falls back to the business‑question matcher.

The question‑matching module applies an end‑to‑end CNN: both the user question and each candidate standard question are embedded into word vectors, passed through shared convolution, pooling, and tanh layers, and compared via cosine similarity to select the most semantically similar answer.

Future plans include adding operational functions (invoice re‑issuance, order cancellation, refund requests) and introducing context‑aware processing to guide users through multi‑turn interactions.

Beyond Qunar, the robot will be integrated into a group‑wide intelligent service platform shared with Ctrip. A unified communication protocol will standardize data collection, enable joint annotation, and allow both companies to maintain a common tourism‑industry knowledge base. The platform will also manage traffic routing, support A/B testing, and provide failover capabilities.

In summary, while the current robot mainly handles informational queries, its iterative technical evolution demonstrates how AI can dramatically improve service efficiency and reduce costs, embodying the philosophy: “Let the machines do the work; humans only think.”