Abstract: To exploit the statistical co-occurrence relationships in line spectra for neural network-based underwater acoustic target recognition, a co-occurrence matrix factorization based line spectrum embedding method is proposed. This paper analyzes the similarity between line spectra, instantaneous spectra, and history spectra in the field of underwater acoustic target recognition and words, sentences, and documents in natural language. Inspired by word embeddings in natural language processing, discrete line spectra are transformed to continuous vectors to represent their statistical co-occurrence relationships using spatial adjacency of vectors, which provides more valuable information for line spectrum analysis. Co-occurrence matrix is used to count the co-occurrence of line spectra and the co-occurrence between line spectra and targets, and matrices are built separately. Matrix factorization is used to transform high-dimensional sparse matrix to low-dimensional real vectors. The generated vectors are applied to underwater acoustic target recognition, achieving a maximum performance improvement of 2.8% compared to the benchmark method.