Data-Driven Concept Extraction for Air Traffic Control Utterances

Kurzfassung

In this thesis, I demonstrated state-of-the-art methods in Natural Language Processing (NLP) to extract the concepts of an ATC utterance. An ATC utterance is speech communication between an Air Traffic Control Officer (ATCO) and Pilot, which follows a standard phraseology. This utterance comprises various elements like callsign, command type, value, unit, conditions, etc. The idea here is to extract the relevant concepts from the recognized word sequence i.e. transcriptions and represent them in the defined ontology format. The thesis mainly investigates an approach to extracting an utterance’s command type, unit, qualifiers, and values. In this framework, the thesis discusses the theory behind state-of-the-art techniques in NLP and Machine Learning (ML). The primary task is divided into text classification and sequence tagging tasks. Furthermore, artificial neural networks like Feed-Forward Neural Network (FNN) and Recurrent Neural Networks (RNNs) are used to predict the command type, unit and qualifier, and value span tagging, respectively. Finally, the selected approach is tested against the Deutsches Zentrum f¨ur Luft- und Raumfahrt (DLR)’s performance metrics. In the instance of gold standard transcriptions, I demonstrated that the chosen technique could obtain a command recognition rate of 94.9% for ATCO utterances and 88.8% for pilot utterances. Nevertheless, the results from the approach selected shows that the NLP techniques could be successfully applied for concept extraction in ATC domain.