<p>The data-driven discovery of luminescent materials is hindered by a critical paradox: while decades of research have generated extensive data, the unstructured textual format of such data in literature precludes its systematic reuse. While text mining enables conversion to structured datasets, general natural language processing (NLP) tools fail at domain-specific challenges, such as fragmented host-dopant syntax and incomplete representations of property relationships. To address these limitations, we developed a novel NLP pipeline specifically tailored to the phosphor domain, integrating three key algorithms: chemically-aware parsing, frequency-driven subject recovery, and context-aware disambiguation algorithms. Using this approach, we successfully extracted 6,400 material-emission relationships from 16,659 scientific papers, with F1 scores of 0.956 for phosphor formula extraction and 0.851 for relation extraction. Additionally, a machine learning model trained on the extracted dataset predicted Eu²⁺ emission wavelengths, with a coefficient of determination (R²) of 0.91. Guided by predictions, we synthesized novel phosphors (e.g., CaGd₂S₄: Eu²⁺), the experimental emission spectra of these materials deviated by around 10 nm from predictions, verifying the robustness of the “data extraction-model prediction-experimental validation” workflow. This study bridges computational discovery and experimental validation, offering an open framework and tools to accelerate luminescent materials innovation.</p>