Classi cation of Readback Errors
Readback errors in English Air Tra c Control speech pose a signi cant safety risk. The National Aeronautics and Space Administration (NASA) and Federal Aviation Administration (FAA) have published various reports illustrating com- munication errors as prominent factors resulting in aviation incidents. While com- munication errors exist in a broad range, readback and hearback errors will be the main focus in this thesis. In an attempt to reduce readback or hearback errors, this research looked at the possibility of automating the hearback process where controllers are required to verify readbacks made by pilots. The task of classifying readbacks as 'Correct', 'Incomplete' or 'Wrong' was seen as analogous to a sentence matching task, where the relationship between instruction from a controller and readback from a pilot was determined. A processing pipeline was introduced as 'The Proposed Scheme' and Natural Language Processing techniques and Convo- lutional Neural Network architectures were explored. The best implementation of the proposed scheme had a 78.2% recall rate, higher than the most recent reported recall rate of human controllers which stood at 77.8%.
Faculteit der Sociale Wetenschappen