Lexicon-Level Contrastive Visual-Grounding Improves Language Modeling