The sense frequency and distribution of Chinese characters commonly used in modern languages of China and Japan have always been a concern of the academic community, as well as a problem in the comparative study of the two languages. Using neural-network-based word embedding as a tool, this paper selects 10 high-frequency homographs with rich meanings as target words to conduct exploratory research on this issue. Research result shows that the sense classification accuracy of Japanese target words based on BERT word embedding reaches 90% on average and 97% at the highest; The average accuracy of Chinese target words has reached 88.3% with the highest also at 97%. It is feasible to use word embedding to carry out quantitative research on the semantics of Chinese and Japanese words. The research also reveals that, among other factors, the scientificity and rationality of word sense induction in traditional dictionaries, the length of dictionary sample sentences, the standardization of corpus sample sentences, the accuracy of sample sentence extraction, and the length of corpus sample sentences have an impact on the semantic analysis based on word embedding.