Due to the difficult of text-to-text semantic similarity feature extraction in spontaneous speech evaluation, this paper presents WordNet based Lesk algorithm to calculate the semantic similarity between words, defines the semantic similarity algorithm between word and text based on the semantic similarity between words, and proposes a complete set of wordnet based text-to-text semantic similarity feature extraction methods. Experiment extracts text-to-text semantic similarity feature between student’s answers and the standard answers with this algorithm and analyzes the correlation between the feature and the teacher rating. Experimental results show that the algorithm can effectively characterize the text-to-text semantic similarity between the students’ answers and the standard answer.