Hi @Li-Fang_Yeo, and welcome to the forum! 
I think this note from the Moving Pictures Tutorial says it all:
Taxonomic classifiers perform best when they are trained based on your specific sample preparation and sequencing parameters, including the primers that were used for amplification and the length of your sequence reads. Therefore in general you should follow the instructions in Training feature classifiers with q2-feature-classifier to train your own taxonomic classifiers. We provide some common classifiers on our data resources page, including Silva-based 16S classifiers, though in the future we may stop providing these in favor of having users train their own classifiers which will be most relevant to their sequence data.
I'm not too sure about the second question, but I think Bod's answer here might be relevant as well: Extracting V4 region from V3-V4 data - #2 by Mehrbod_Estaki.