I guess there is no easy answer, because the Tagger is trained with a statistical model (I found it being trained with a Back-off Trigram Markov Model on the Penn Treebank here).
I could imagine "select icon from icon" being a very rare occurrence in the training corpus (if it occurred at all), so looking at the first word, having no more contextual information except maybe for the beginning of a new sentence, it assigned JJ
as the most likely tag.
If this is a huge problem for you, you can consider training your own tagger on a corpus where more of these kinds of sentences occur or enrich an old one using something like this.