AI‑Driven Intelligent Customer Service at Ctrip: Algorithms and Practices

Overview

Ctrip, the leading domestic OTA, serves millions of travelers daily, and its massive customer‑service operation contains a large amount of repetitive work. By applying recent breakthroughs in deep‑learning and upgraded hardware, Ctrip has built an intelligent customer‑service system that currently resolves about 70% of hotel‑after‑sale queries, significantly boosting efficiency and response speed.

Question‑Answer Match

The basic QA matching model starts from keyword or BM25 retrieval, then progresses to latent semantic analysis (LSA), probabilistic LSA, LDA, and word2vec. Because word2vec alone cannot capture phrase or sentence semantics, Ctrip adopts sentence‑level neural models such as DSSM, MV‑LSTM, MatchPyramid, and other CNN/LSTM‑based architectures, typically built with fully‑connected layers, LSTM, convolution, and pooling.

Two training paradigms are used:

Classification : a multi‑class softmax predicts the standard question category.

Ranking : point‑wise (binary relevance) and pair‑wise (relative ordering) losses rank candidate questions. The point‑wise structure and pair‑wise structure are illustrated in the original figures. Single‑Turn vs. Contextual Dialogue Single‑turn models cannot handle follow‑up queries, so Ctrip incorporates multi‑turn context modeling. Two approaches are discussed: Rule‑based context representation using predefined slots and states. Model‑based context representation that encodes utterances into vectors with hierarchical encoders (sentence encoder, context encoder, response decoder). The model architecture is shown in the accompanying diagram. Input Suggestion Input suggestion helps users formulate queries with semantic similarity rather than exact keyword matches. Ctrip iterated three algorithms: Version 1: a simple Trie tree for prefix matching, later enhanced with stop‑word removal. Version 2: a point‑wise ranking model using logistic regression with features from word2vec, TF‑IDF, and domain‑specific keywords, which greatly improved recall. Version 3: Microsoft’s DSSM with word‑hashing to handle OOV and short texts, delivering the best performance. Each version’s impact is shown in the original performance charts. Conclusion Beyond the scenarios described, various machine‑learning and deep‑learning algorithms assist Ctrip’s customer‑service workflow, enhancing efficiency and user experience. The article emphasizes that AI is not a mysterious black box but a combination of data, algorithms, and human expertise, requiring engineers to integrate all three components into a closed loop.