From bursty patterns to bursty facts: The effectiveness of temporal text mining for news