External text can provide predictive signals that are absent from pure numerical histories, yet its usefulness hinges on how the text is represented, combined, and deployed. We introduce Attribution‑Guided Semantic Evolution (ASE), a modular framework for constructing textual factors, evaluating their conditional contributions, and deploying validation‑selected numerical‑textual predictors.\ \ ASE supports task‑specific implementations of three kinds of textual factors: lexical measurements, semantic primitives, and evolved event/state factors, each coupled with appropriate numerical expert models. The feedback‑driven implementation uses conditional refitting diagnostics to steer factor revision and combination search, dynamically adjusting the weight and form of textual factors at each iteration.\ \ We evaluate ASE on the four text‑associated TESS datasets and a dynamic CSI 300 stock ranking task. The four forecasting implementations achieve test MSEs of 0.001041, 0.022439, 0.333115 and 1.8150, each outperforming the corresponding published TESS scores. On CSI 300, the frozen text‑fusion implementation improves mean RankIC by 0.000642 over the numerical‑only expert and by 0.000701 over a matched shuffled‑text baseline, with positive gains across all five random seeds.\ \ Controlled text perturbation experiments confirm the model’s functional reliance on text correspondence, while an incumbent‑retention mechanism prevents regressions in the measured development criterion. Together, the experiments characterize ASE at two levels: forecasting performance of task‑specific implementations and the conditional contributions of individual modules.\ \ Review