Natural Language to Code: Improving Semantic Reasoning in Code Generation Models

Creating code from human-readable instructions is becoming a major area of research as artificial intelligence is used more and more into software engineering procedures. This paper explores techniques to enhance semantic understanding in AI-based code generation models to improve their ability to interpret human intent and produce accurate, executable code. We investigate the performance of state-of-the-art models such as CodeT5 and PLBART, and propose strategies including prompt engineering, domain-specific fine-tuning and execution-aware evaluation metrics. Our experiments are conducted on datasets like MBPP and APPS, where we evaluate both syntactic correctness and functional accuracy of generated code. Results show that incorporating contextual awareness and structured prompting significantly improves code quality and reduces semantic misinterpretation errors. The findings contribute to the ongoing effort to build more intelligent, reliable and context-aware coding assistants.