Exploring the Role of Named Entities for Uncertainty Recognition in Event Detection

摘要

Ambiguous information contributes to the uncertainty issue. Type of information such as using named entities has been proved to provide significant information to the user compared to the 'bag-of-words' in identifying an event. So what else could contribute to the uncertainty in an event detection? We propose to answer this question by analysing the distribution of named entities across topics, and explore the potential of named entities in a user experiment. We construct an event detection task with 20 users and use news dataset from Topic Detection and Tracking (TDT) corpus, under the Sports and Politics categories. We analyse the results from five uncertainty dimensions: too little information, too much information, complex information, ambiguous information and conflicting information. These dimensions are categorise as two factors; amount and type of information. There was no statistical significance difference in the amount of information given with the number of successful event detected. However, with little information and high named entities has contributes in reducing uncertainty. In addition, the amount of information and information quality are mutually independent. Our results suggest that uncertainty vary substantially between the amount of information and type of information in event detection.