Deep Learning Applications in Ctrip Travel Guide Community

Editor: This article is based on a keynote speech by Li Jian, Development Director of Ctrip Strategy Community, presented at the Ctrip Technology Center Deep Learning Meetup. The original PPT can be downloaded from the source.

Ctrip Strategy Community, a travel‑guide platform under Ctrip, serves millions of travelers with itineraries, reviews, and travel notes. Backed by 250 million Ctrip users, the community has over 5 million daily active users, 30 million authentic reviews, 400 000 travel logs, and 2 000 influential travel experts. Faced with massive, heterogeneous data, the challenge is to extract valuable information efficiently.

Main Requirements of the Community

The data is complex: multilingual textual content ranging from short comments to long travelogues, official attraction pages, and user‑uploaded images. Models must quickly and accurately classify this massive, varied information.

Brief Introduction to Deep Learning and Classic Models

The talk contrasts shallow machine‑learning models with deep learning. Starting from traditional Support Vector Machines (SVM), it progresses to Convolutional Neural Networks (CNN), which play a pivotal role in the presented solutions.

Key characteristics, advantages, and application scenarios of CNN are explained. In addition to CNN, the presentation covers Word2Vec, Recurrent Neural Networks (RNN), Long Short‑Term Memory networks (LSTM), and Recurrent Convolutional Neural Networks (RCNN).

Deep Learning in the Strategy Community

After outlining the community’s specific needs and the mainstream deep‑learning models, the speaker describes how to match models to tasks efficiently.

1. Natural Language Processing : CNN for sentiment analysis, LSTM for address‑quality scoring, and a combined CNN‑Highway‑LSTM‑Attention model for extracting and judging opening times of specific attractions.

2. Image Processing : CNN for image classification to detect advertisements, infringing content, and illegal images; CNN‑LSTM‑Attention for object detection and automatic generation of sentiment‑rich captions.

3. Video Domain : RCNN and LSTM models for automatic video moderation and generation of corresponding textual descriptions.

4. Data Content Matching : Fuzzy neural networks for multi‑dimensional matching of destinations and Points‑of‑Interest (POI) information.

Future Trends and Focus Points

Going forward, the community aims to optimize existing models and explore automatic error correction, knowledge graphs, virtual reality, and broader deep‑learning architectures to improve computational efficiency, accuracy, and user experience.

Recommended Reading

Building a Technical Sharing Platform – Review of Ctrip Technology Center Deep Learning Meetup