Efficient methods for topic model inference on streaming document collections