Common forms of lexicalization, acronyms are characterized by their concise structure, precise meaning, and retention of the original intent. They can effectively condense text and convey key information, playing a crucial role in various natural language processing tasks such as information retrieval and entity linking. However, the structure of Chinese acronyms is complex and diverse, and some acronyms are polysemous, which means they may represent different entities in different contexts. Most existing methods primarily rely on the entities themselves and rarely incorporate contextual information when studying Chinese acronyms, thereby overlooking the rich semantics within the context. Therefore, this paper first enhances a large Chinese acronym dataset with contextual information. Then, it proposes a new generation-evaluation framework for Chinese acronym prediction. This framework consists of a generation model that generates multiple candidate acronyms and an evaluation model based on thematic consistency. Experimental results on the public dataset demonstrate the effectiveness of our method, achieving improvements of 3.5 and 1.6 in Hit@1 and Hit@3, respectively.