Abstract: A previous text steganography scheme uses the mimic function and the Context-Free Grammar (CFG) to generate pseudo-natural language text to convey secret information. However, the generated text usually lacks semantic consistency and is therefore vulnerable to human evaluation. An improved mimicry technique is proposed that constructs the CFG-based grammar file from a template text that meets certain criteria so that the generated text is semantically consistent. The grammar file is constructed from a replacement text base composed of many optional words, phrases or sentences. Huffman coding is used to take into account the occurrence frequencies of various words, expressions, and syntactic structures to improve naturalness of the generated text. Thus, in addition to machine examination, the obtained stego-text is able to pass stringent human evaluation.