Using Domain Knowledge about Medications to Correct Recognition Errors in Medical Report Creation

We present an approach to analysing automatic speech recognition (ASR) hypotheses for dictated medical reports based on background knowledge. Our application area is prescriptions of medications, which are a frequent source of misrecognitions: In a sample report corpus, we found that about 40% of the active substances or trade names and dosages were recognized incorrectly. In about 25% of these errors, the correct string of words was contained in the word graph. We have built a knowledge base of medications based on information contained in the Unified Medical Language System (UMLS), consisting of trade names, active substances, strengths and dosages. From this, we generate a variety of linguistic realizations for prescriptions. Whenever an inconsistency in a prescription is encountered on the best path of the word graph, the system searches for alternative paths which contain valid linguistic realizations of prescriptions consistent with the knowledge base. If such a path exists, a new concept edge with a better score is added to the word graph, resulting in a higher plausibility for this reading. The concept edge can be used for rescoring the word graph to obtain a new best path. A preliminary evaluation led to encouraging results: in nearly half of the cases where the word graph contained the correct variant, the correction was successful.