⁶In Chapter 6, I am already dealing with a very noisy environment. Verbalizing the knowledge lets me gain more information from the training data as the semantic classes are too general. For example, generating the text and comparing it to the target text is necessary to distinguish tachycardia from bradycardia problems (both are mapped to the “intra-op-problems” semantic tag). That is not the case with the matched texts presented in this chapter.