Category-Level transfer learning from knowledge base to microblog stream for accurate event detection

Abstract

Many Web applications need the accurate event detection technique on microblog stream. But the accuracy of existing methods is still challenged by microblog’s short length and high noise. We develop a novel category-level transfer learning method TransDetector to deal with the task. TransDetector bases on two facts, that microblog is short but can be enriched by knowledge base semantically with transfer learning; and events can be detected more accurately on microblogs with richer semantics. The following contributions are made in TransDetector. (1) We propose a structure-guided category-level topics extraction method, which exploits the knowledge base’s hierarchical structure to extract categories’ highly correlated topics. (2)We develop a probabilistic model CTrans-LDA for category-level transfer learning, which utilizes the word co-occurrences and transfers the knowledge base’s category-level topics into microblogs. (3) Events are detected accurately on categorylevel word time series, due to richer semantics and less noise. (4) Experiment verifies the quality of category-level topics extracted from knowledge base, and the further study on the benchmark Edinburgh twitter corpus validates the effectiveness of our proposed transfer learning method for event detection. TransDetector achieves high accuracy, promoting the precision by 9% without sacrificing the recall rate.