Hotspot Mining and Event Extraction in Tencent Information Flow: Methods, Framework, and Applications

Introduction – Modern news and social media platforms prominently display hot topics, and the ability to quickly discover, organize, and personalize these hotspots is crucial for user experience.

Project Background – The authors identify four key problems: (1) timeliness – rapid detection of events such as Kobe Bryant’s death; (2) comprehensiveness – aggregating diverse reports across sites; (3) heat rationality – differing popularity across data sources; and (4) distribution – handling cold‑start issues for newly emerging events.

Related Research Methods

Event Extraction – Detecting trigger words and classifying event types (e.g., "die" for Kobe’s death) using ACE‑style frameworks.

Topic Detection & Tracking (TDT) – Clustering queries and documents to form topics, employing k‑means, DBSCAN, hierarchical clustering, and enriched semantic, entity, and event features.

Dynamic Time Warping (DTW) – Aligning time‑series of query frequencies to templates for hotspot identification.

Hotspot Computation Framework – Consists of offline mining and online understanding. Offline mining ingests data from three Tencent platforms, performs topic extraction, topic fusion, and event splitting; online understanding matches new content to the event library.

Hotspot Mining Pipeline

Query‑Log Mining – Build time‑series from search queries, detect popular queries using DTW, and cluster similar queries into topics.

Article Mining – Convert article content into keyword features, apply DTW‑based hotspot detection, then perform hierarchical clustering to form topics while filtering non‑hot content.

Topic Fusion – Merge topics from different sources, compute a unified heat score combining consumption and production signals.

Event Splitting – When a topic reaches sufficient granularity, split it into distinct events using trigger‑argument analysis and seq2seq‑based naming.

Heat Calculation – Combine consumption heat (user searches), production heat (author activity), and global heat to assign final scores for recommendation.

Applications

Image‑Text Hotspot – Use a dual‑tower and MatchPyramid model to match article titles with event names, assigning event and topic tags for recommendation.

Video & Short‑Video Hotspot – Transfer textual hotspot signals to video content, filter by popular keywords, and match video titles to event names for personalized feeds.

Conclusion – The presented framework improves hotspot relevance, timeliness, and personalization across multiple media formats, supporting Tencent’s information‑flow recommendation system.