Incremental Novelty Learning in Adaptive Topic Tracking

As an important task in Topic Detection and Tracking (TDT), Topic tracking aims to monitor story stream arranged in temporal order and identify relevant stories to given topics. On this basis Adaptive Topic Tracking focuses on modifying topic descriptions by learning feedbacks in real time to improve the accuracy of tracking. Many researches have achieved substantial performances in this field, however most of them attempt to resolve the issues of tracking in use of existing techniques in Information Retrieval which often ignore the characteristics of TDT corpus such as timeliness, novelty and burst of news stories. In this paper we propose an incremental novelty learning algorithm (INL) which learns the evolution of topic based on the incremental distribution of novel terms and modifies the topic description by combining seminal terms with novel ones in real time. The algorithm focuses on increasing the efficiency of updating the kernel of topic description and improving the ability of tracking novel events suddenly occurred in story stream. We compare our method with some existing tracking systems on TDT4 corpus, which demonstrates INL can substantially improve the accuracy of adaptive topic tracking.