HU_DB at TREC 2014 Microblog Track

This paper describes our system for the Tweet Timeline Generation (TTG) task of the Microblog track, at the Text Retrieval Conference (TREC) 2014. Intuitively, given a collection of microblog posts (i.e., tweets), and a keyword query Q, the goal is to generate a timeline of related tweets. Such a timeline consists of representative tweets, relevant to Q. In our system we employ query expansion to identify highly relevant tweets, and then use affinity propagation to cluster the tweets, based on their word similarity, hashtag similarity and temporal similarity. We then return a representative tweet from each cluster. The result is a system with relatively good precision, but, unfortunately, poor recall. We discuss the techniques employed, as well as the insights gleaned while developing and testing our system.