<p>Keyphrases are concise semantic units that capture the meanings of a document, providing interpretable abstractions that reduce textual complexity and support downstream applications such as information retrieval, text summarization, and document classification. While supervised keyphrase prediction methods rely on costly annotated data, unsupervised methods have emerged as effective alternatives. This survey provides a comprehensive overview of unsupervised keyphrase prediction methods and systematically reviews the evaluation metrics. We analyze the linguistic properties of keyphrases to inform the design of evaluation metrics and guide the improvement of prediction methods. In addition, we categorize existing unsupervised approaches and discuss their respective strengths and limitations. Furthermore, we review existing evaluation metrics and analyze their shortcomings. To address these gaps, we introduce CAME, a novel reference-free metric, and validate its effectiveness through experiments. Finally, we highlight challenges and promising directions, particularly in the context of LLMs, to guide future research in unsupervised keyphrase prediction.</p>