Bigger versus Similar: Selecting a Background Corpus for First Story Detection based on Distributional Similarity