Classification of Air Traffic Controller Utterance Transcripts via Warm-Start Non-Negative Matrix Factorization

Air traffic control voice (i.e., utterance) transcript data are often underutilized in the context of airspace analysis, despite its increasing availability. This underuse provides an opportunity for enhanced analysis, as utterances contain operational directives which are typically inferred from aircraft trajectories. Direct knowledge of such directives would facilitate various activities such as Area Navigation (RNAV) procedure assessment, airspace redesign, or controller workload studies. Though transcribed utterances can be free-form, they fit into a finite number of categories. To this end, we propose using domain knowledge to create an effective warm-start strategy for the non-negative matrix factorization (NMF), which in turn can be used to categorize utterance transcripts automatically. Using human and machine transcribed voice data, we show that our approach closely matches manually labeled (i.e., by subject matter experts) utterances. Furthermore, we associate labeled utterances to their corresp...