Most of the existing research on Physical-layer Network Coding (PNC) is based on the assumption that the symbol timing at the relay is ideally synchronized, and rarely discusses the issue of symbol synchronization. However, in practice, the symbol timing is indispensable in PNC systems. To tackle this problem, this paper proposes a novel symbol timing estimation scheme based on the orthogonal training sequences for PNC in two-way relay channels. According to the maximum-likelihood estimation criterion, a Discrete Fourier Transformation (DFT) based interpolation algorithm is applied to improve the estimation accuracy. It is shown by analysis and simulation that the proposed DFT-based symbol timing estimator exhibits superior performance. The Mean Square Error (MSE) performance of the estimator is one order of magnitude better than that of the conventional optimum sample algorithm for Signal-to-Noise Ratio (SNR) greater than 10 dB, and is very close to the Modified Cramer-Rao Bound (MCRB).