Real-time filtering on interest profiles in Twitter stream

The advent of Twitter has led to the ubiquitous information overload problem with a dramatic increase in the amount of tweets a user is exposed to. In this paper, we consider real-time tweet filtering with respect to users' interest profiles in public Twitter stream. While traditional filtering methods mainly focus on judging relevance of a document, we aim to retrieve relevant and novel documents to address the high redundancy of tweets. An unsupervised approach is proposed to model relevance between tweets and different profiles adaptively and a neural network language model is employed to learn semantic representation for tweets. Experiments on TREC 2015 dataset demonstrate the effectiveness of the proposed approach.