Uncovering Limitations in Text-to-Image Generation: A Contrastive Approach with Structured Semantic Alignment F i n d i n g s - E M N L P