The Best of Both Worlds: Context-Powered Word Embedding Combinations for Longitudinal Text Analysis