Unraveling the Dynamics of Semi-Supervised Hate Speech Detection: The Impact of Unlabeled Data Characteristics and Pseudo-Labeling Strategies F i n d i n g s - E A C L