Estimation of dielectric constants is crucial for locating underground cavities. A typical method for determining the dielectric constants of soil layers is the inverse analysis based on ground penetrating radar (GPR) data. This study proposes an efficient probabilistic inverse analysis method for multi-layer dielectric constants by integrating Bayesian inference and a Bidirectional Long Short-Term Memory (BiLSTM) model. The BiLSTM model serves as the forward model in the inverse analysis, which is trained based on a dataset generated from finite-difference time-domain (FDTD) simulations. The dielectric constants and thicknesses of soil layers are inferred by incorporating measured data, and their posterior distributions are obtained through Markov Chain Monte Carlo (MCMC) sampling. A three-layer soil case is used to illustrate and validate the proposed method. The results show that the proposed method can efficiently and accurately estimate the dielectric constants and thicknesses of multi-layer soils, while also effectively quantifying the associated uncertainties.