Short Text Understanding in the Medical Health Domain: Business Scenarios and Technical Practices

The talk introduces short‑text understanding in the medical‑health domain, focusing on DingXiangYuan’s business scenarios and the challenges of processing both professional and consumer medical texts.

It describes various search applications such as the DingXiangYuan forum, drug assistant, doctor app, and e‑commerce platform, highlighting the need to handle topic‑driven, complex, and serious medical content.

The NLP pipeline is divided into five steps: text correction, noun‑phrase extraction, named‑entity recognition, mention extraction, and entity linking, with details on spell checking, candidate generation, CRF+Bi‑LSTM NER, and ranking methods.

Knowledge‑graph construction and content profiling are employed to enrich entities and generate abstract concept tags, using graph embeddings and MDL‑based filtering to ensure high‑quality concepts.

Extracted concepts enable downstream tasks such as query expansion, tag generation, and recommendation in both professional and consumer scenarios, improving relevance and coverage.

The conclusion emphasizes improving entity‑recognition accuracy, mining concepts from user logs, and building a business graph to support search and recommendation pipelines.