This paper presents a semantic analysis of corpus-assisted collocation teaching based on the random matrix theory, taking into account the characteristics of the computer corpus. The construction of a word-document matrix is based on the potential semantic analysis of the corpus and the intrinsic relationships between words and documents. The aggregation and combination relations of collocation, i.e., the lexicon of word meanings, are used to calculate affirmative sentence similarity, incorporating dependent syntactic and semantic analysis. The error analysis method was applied to the teaching of collocations to analyze the common types of errors and the variability of writing errors in English writing and to discuss the correlation between the performance of English composition and the number of collocations used. The correlation coefficient between composition scores and the number of collocations in compositions was 0.924, and the composition scores of college students had a positive linear relationship with the number of collocations.