A study on the joint channel and symbol estimation issue is provided for two hop relay networks which employ the amplify-and-forward (AF) relaying approach.The encoding scheme at the source node introduces the time-domain spreading with a time-varying linear constellation precoding.Then,a set of amplifying factors matrices is utilized by the relays to amplify and forward the received data to the destination.The received signal at the destination can be constructed as a fourth-order tensor model,which is referred to as the nested parallel factor (PARAFAC) model.And then,we present a novel Levenberg-Marquardt (LM) algorithm based on this tensor model.The proposed method does not require complex signal processing at the relay,which effectively reduces the burden of relay.As a semi-blind method,which does not require the pilot signal,the proposed receiver can jointly recover the channels and information symbols.Moreover,the proposed semi-blind receiver is robust as it can work in different wireless channel scenarios.Simulations are conducted to demonstrate the efficiency of the proposed semi-blind approach.