Assessment and manipulation of latent constructs in pre-trained language models using psychometric scales