Design and Implementation of a One‑Stop Customer Service Chatbot at Ctrip

Author Bio Jingwei works on customer‑service robot algorithm research at Ctrip, focusing on NLP algorithms and data‑driven development.

Ctrip, as a comprehensive OTA, offers many business lines (flights, hotels, trains, tickets, car rentals, etc.). Currently, each line has its own customer‑service channel, forcing users to locate the appropriate line before consulting, which is sub‑optimal for a full‑travel experience. Figure 1 shows the existing multi‑line entry process.

Figure 1 Pre‑sale consultation entry for each business line

To provide a true "one‑stop" experience, Ctrip plans to launch a unified chatbot in its app that allows users to resolve all travel‑related issues from a single entry point.

The article first compares the current multi‑line entry with the proposed one‑stop entry, highlighting the importance and challenges of building such a chatbot, then presents the complete system architecture and details the natural‑language‑understanding (NLU) component.

1. Comparison of Existing Multi‑Line and One‑Stop Entry

Example 1 (Figure 2) shows a user who cancelled a trip and needs refunds for both flight and hotel. In the current flow, the user must switch from the flight‑refund channel to the hotel‑refund channel, because each channel only knows its own domain. In the one‑stop entry, the user can complete the entire process without switching.

Figure 2 Flight order page vs. one‑stop chatbot entry

Example 2 (Figure 3) illustrates a hotel‑information and recommendation scenario. The hotel‑detail page can only answer questions about that specific hotel; recommendation requests are transferred to a human. The one‑stop chatbot can both answer hotel‑specific queries and provide nearby hotel recommendations.

Figure 3 Hotel detail page vs. one‑stop chatbot entry

In summary, the one‑stop chatbot offers two main advantages: (1) it possesses knowledge across all business lines, providing comprehensive travel support; (2) it supports additional skills such as recommendation and booking, improving user convenience and enabling precise marketing.

2. One‑Stop Chatbot Architecture

The architecture (Figure 4) consists of an NLU module (domain classification, intent classification, slot filling), a dialogue‑management module that updates conversation state and selects appropriate actions, and a response‑generation module that invokes Q&A, recommendation, or other skill services.

Figure 4 Algorithm‑side architecture of the one‑stop chatbot

Two concrete scenarios (Figures 5 and 6) demonstrate how the system handles cross‑line refund queries and hotel‑facility recommendation by leveraging NLU outputs and dialogue management to invoke the correct FAQ or knowledge‑base services.

Figure 5 Cross‑business‑line service fulfillment

Figure 6 Hotel query and recommendation flow

3. Natural Language Understanding Algorithms

The NLU module performs domain classification, intent classification, and slot filling. Intent classification is cascaded: first predict the domain, then use the domain embedding to aid intent prediction. Input tokens are one‑hot encoded, embedded (character‑level, word‑level, or hybrid via CNN/Highway), then encoded by LSTM/GRU. A fully‑connected layer with softmax yields domain predictions; the domain embedding is combined with sequence representations for intent classification, again using LSTM/GRU and attention.

Figure 7 Domain and intent classification architecture

Experiments show that domain‑aware attention improves classification accuracy by ~1.5‑2 %, and the cascaded intent classification approach adds another 2‑3 % gain compared with flat classification. Offline training achieved ~90 % accuracy for domain classification and ~87 % for intent classification, meeting production standards.

4. Conclusion

The paper presented the vision, benefits, and challenges of Ctrip’s one‑stop chatbot, described its algorithmic architecture, and detailed the NLU component that enables cross‑line query handling. Future work includes finer‑grained NLU models, more accurate intent and slot prediction, fault‑tolerant dialogue management, richer knowledge bases, and extensible intent handling.