In this investigation, we propose a new method to estimate headlines to news articles. Very often, in news articles, headlines contain characteristic expressions specific to their contents. However, conventional approaches may extract keywords or patterns from article bodies, and put them into well-forms. However we can hardly obtain the characteristic expressions. Here we examine both news articles and the headlines separetely, give bridge between the two documents in such a way that similar articles carry similar headlines. To do that we examine latent semantics over articles to decide similarity, and generate headline candidates semi-automatically.