Intelligent Customer Service in Travel: System Architecture and Key Technologies

Author Introduction

Lei Lei, senior algorithm engineer in the Ctrip Vacation R&D department, is responsible for intelligent customer service algorithms. Ju Jianxun, algorithm manager in the same department, works on intelligent customer service, knowledge graph, and NLP algorithms.

With the development of artificial intelligence, human‑computer interaction technologies have matured, and intelligent customer service has become a key application in the travel domain, serving both customers and support staff.

The current Ctrip intelligent customer service serves two roles: a client‑side bot for travelers and an assistant for customer service agents.

For agents, the assistant appears in a sidebar, suggesting possible answers; for travelers, the bot directly answers queries.

Intelligent customer service includes single‑turn QA bots and multi‑turn Task bots; in travel scenarios, Task bots dominate. A typical multi‑turn system processes a user query through NLU to extract intent and slots, passes them to a dialogue manager (DM), then through DST, DPL, and NLG modules to generate a response.

NLU (Natural Language Understanding)

NLU extracts user intent and slot information using models or rules. In complex travel scenarios, multi‑turn Task bots are essential.

1.1 Spell Correction

Before intent detection, a correction module fixes typographical errors using a rule‑based component for common place names and a CRF model followed by a language model for remaining errors.

1.2 Intent Recognition

Intent abstracts a user question (e.g., “How much does this product cost?” → “price inquiry”). A weak‑supervision matching‑network model maps queries to intent vectors and selects the most similar intent.

Training uses C known intents with N examples each; sentences are encoded by Bi‑LSTM and Transformer‑Attention, then classified via a capsule network.

1.3 Handling Fuzzy Intents

When users input short or ambiguous queries, “suggested questions” and “you may ask” features retrieve similar frequently asked questions using text similarity algorithms.

1.4 Discovering New Intents

New business lines gradually add intents. Hierarchical clustering (Gaussian Mixture → OPTICS) groups similar queries to propose new intents for business review.

1.5 Slot Extraction

Slot extraction supports answer retrieval (e.g., visa requirements). A hybrid rule‑model approach covers ~70% of cases; the remaining are handled by a BERT‑based Bi‑LSTM‑CRF pipeline.

2. Dialogue Management System

The dialogue manager handles DST (state tracking), DPL (policy selection), and NLG (response generation). It receives intents and slots from NLU, updates the dialogue state, and decides the next action (answer, clarification, etc.).

3. Intelligent Customer Service Platform

The platform decouples business configuration from technical implementation. Users create projects, define intents, slots, and answer templates, and can configure answers as static text, knowledge‑graph schemas, external APIs, or slot‑based responses.

4. Conclusion and Future Work

The presented technologies have been deployed in a complete intelligent customer service solution. Future directions include multi‑modal (voice, image) and multilingual support, proactive service modes, human‑AI collaboration, group chat capabilities, and large‑scale data mining to reduce annotation costs.

Recommended Reading

One‑click conversion of WeChat Mini‑Program to Baidu Mini‑Program by Ctrip Train Ticket Team

Ctrip’s Journey with Dubbo

Improving Data Quality at Scale

Ctrip Redis Cross‑IDC Multi‑Direction Sync Practice

2019 Ctrip Technology Summit – November 9, Shanghai. Use the public account discount code “ctriptech” for an 20% ticket discount.