Content area
Abstract
Anomaly detection from Aircraft Safety Reporting System (ASRS) reports is the task of identifying the multiple types of anomalies that resulted in an aviation incident based on the narrative portion of the report, which was written with lot of domain specific terminology. The thesis proposes an effective semi-supervised approach to solving the problem without using any domain related information. The approach relies on methods similar to singular value decomposition and latent semantics for feature generation. A Support Vector Machine with transduction was used to classify the narratives. Experimental results show that the method developed was marginally better than the existing methods for high frequency anomalies and substantially better for low frequency anomaly on the ASRS dataset.
