Applying Snorkel Weak Supervision to Automate Event Summaries in Ctrip Customer Service

In recent years, the rapid development of deep learning has highlighted the critical need for large, high‑quality labeled datasets, yet manual annotation is costly, time‑consuming, and often infeasible for specialized domains such as hotel customer service.

The article introduces Snorkel, a weak‑supervision system that allows developers to write labeling functions (LFs) – small pieces of code that assign noisy labels based on heuristics, patterns, external resources, or pre‑trained models – to automatically generate training data without exhaustive manual effort.

For Ctrip’s “event summary” scenario, the workflow consists of data collection, exploratory analysis, LF design (emphasizing precision over recall), LF filtering with a majority‑vote baseline, training a generative label model to combine LF votes, and finally training a discriminative BERT‑based classifier on the generated labels.

The generative model improves coverage and accuracy over the baseline voting model, while the discriminative model further boosts performance, especially for the minority class (hotel refusal), achieving notable gains in both precision and recall.

Future work includes exploring Snorkel’s transformation functions (TF) for data augmentation and slice functions (SF) for targeted modeling, aiming to further enhance model robustness in production.

The article concludes with a brief team recruitment notice, inviting engineers in backend, frontend, mobile, data, and algorithm roles to join Ctrip’s hotel R&D team.