Radar word is a fixed arrangement of finite number of pulses that can reflect the work modes and threat levels of hostile multi-function radars. To deal with the problem of radar words extraction with reference to dropped pulses and false pulses, this paper proposes a new extracting method based on three-level matching (TLM) algorithm, which can first perform a coarse search by identifying the radars from the radar template library that have character parameters similar to those present in the deinterleaved pulse stream of the emitter, and then make use of the technology of pulse pairing and sequence cross correlation for precise identification. A number of simulations are presented to demonstrate the radar words extracting capability of the TLM algorithm. Simulation results show that the proposed algorithm can not only extract the radar words correctly, but also have better adaptability to noisy environment.