<p>Mental health disorders pose a significant global challenge, motivating growing interest in natural language processing (NLP) methods for automated mental health assessment. In recent years, the field has evolved rapidly from traditional feature-based approaches to deep learning architectures and pre-trained foundation models. However, a comprehensive understanding of their relative strengths, limitations, and practical implications remains limited. This paper presents a survey of NLP methodologies for mental health assessment, covering commonly used data sources and representative modeling approaches, and analyzing how advances in representation learning have influenced assessment capability, interpretability, and deployment feasibility. Given that existing studies often rely on disparate datasets and metrics, a direct comparison of these methodologies remains difficult. To complement the literature synthesis, we conduct a unified empirical comparison of representative methods under a consistent experimental setting, providing an additional perspective on performance and efficiency trade-offs. Based on both the literature survey and empirical observations, we discuss key insights that shape the practical use of NLP in mental health, including trade-offs between model complexity and scalability, the role of instruction adherence in prompting-based reasoning, and persistent limitations of current datasets and benchmarks. Building on these observations, this survey outlines important challenges and future research directions toward the responsible and scalable application of NLP technologies in mental health assessment.</p>