[Purpose/significance] To address the limitations of existing disruptive technology identification methods,which primarily rely on static analysis and struggle to continuously characterize dynamic evolution or accurately quantify the direction of trajectory shifts and breakout rates, this study constructs a disruptive singularity identification computational model integrating Natural Language Processing (NLP) and geometric dynamics, aiming to provide quantitative intelligence support for technological foresight and strategic decision-making. [Method/process] Utilizing the pretrained SciBERT model and sliding window techniques, discrete patent texts are mapped into a continuous, dynamic semantic vector space to extract the technological centroid. Drawing upon physical kinematics theory, the tangential acceleration and normal evolutionary curvature of the centroid's trajectory are calculated to represent the short-term concentration of R&D resources and technological route mutations, respectively. Furthermore, by combining Kernel Density Estimation (KDE) with the long-tail characteristics of the data, an identification mechanism for disruptive singularities is established using the superposition of dual statistical extremes (Top10%) as the strict criterion. [Result/conclusion] Empirical research in the gene-editing domain demonstrates that this model successfully captured the disuptive singularities associated with three transformative paradigm shifts. Compared to traditional patent volume mutation models, it can identify key signals 2-3 years in advance; compared to static semantic novelty models, it effectively filters out the pseudo-mutation noise of conventional incremental innovations. This study achieves a methodological expansion in technological evolution analysis from static scientometrics to dynamic kinematic characterization, providing robust intelligence support for technological foresight and strategic decision-making.