Visual-Semantic Alignment Across Domains Using a Semi-Supervised Approach