btw i strongly suggest working further on image analysis so the algorithm takes not only textual elements as an input but visual ones as well.