Traditional lossless text compression preserves every byte, but its gains on natural language are often modest. We study lossy semantic text compression, where the encoder strategically deletes parts of the text and a large language model (LLM) reconstructs the original from the retained skeleton. We benchmark a progression of deletion strategies, from character-level and word-level rules driven by word length or corpus frequency to surprisal-based and hybrid rules that require a neural encoder, and reach three findings. First, deletion by corpus word frequency is a strong low-cost baseline: a static frequency lookup remains competitive with far more expensive semantic methods. Second, semantic and hybrid methods gain most at mild-to-moderate compression, whereas at the lowest retention rate frequency-based deletion matches them and keeps more named entities. Third, QLoRA fine-tuning yields a compact local decoder that, once adapted, is competitive with zero-shot Gemini 2.0 Flash. Additional English and Chinese experiments show that the framework transfers across domains, while the best deletion rule remains dataset-dependent.