In this paper, we consider the problem of channel estimation for two-way relay networks (TWRN) under time-selective environment. We first parameterize the time-varying channels by the basis expansion model (BEM) and then propose a novel pilot symbol aided modulation (PSAM) for TWRN. A linear approach to estimate the cascaded channels is designed and the optimal training sequences are derived based on minimizing the mean-square error (MSE) criterion. Moreover, we develop an algorithm to recover the individual channel knowledge with which both the channel estimation accuracy and the system performance can be improved. Various simulations are provided to corroborate the proposed studies.