The replacement of textual units by synonymous canonical forms is an important prerequisite for many variants of automated text analysis. In scientific texts, one common normalization step is the consistent replacement of acronyms by their definitions. For many acronyms, the definition is found at a certain position of the text where the acronym is introduced and “expanded” to a synonymous sequence of full words. A recent approach to detecting acronym-expansion pairs by Park and Byrd [19] describes possible graphical correspondences between acronyms and expansions by means of fine-grained rules. Here we show how rule sets as used in [19] can be translated into hidden Markov models that abstract from details of the graphical correspondence and improve recall in a significant way. Stability in terms of precision is ensured by exploiting simple properties of the expansion with an optional reinforcement of linguistic knowledge. With this extension of the original formalism, the introduction of large rule sets can be avoided and a fixed model can be applied to a large variety of texts without retraining, with good values both for recall and precision.