Applying Semantic Text Similarity to support the Identification of Human Factors in Aviation Accidents

In Aviation Safety Management, the final goal of the investigator, when approaching an accident, is to identify which are those factors that have realistically contributed to causing the unfortunate events leading to the accident. The aim of investigations is to draft, starting from the causes of the accident, safety recommendations and remedial actions that can eliminate avoidable human, economic and social costs. Previous work was conducted to formally describe how the analysis of an aircraft incident/accident could be leaded in a partially automatic way, developing a supporting expert system that can address the safety investigator during his inquiry. Currently, the system is able to process accident reports and classify the events following a particular safety standard. Purpose of this thesis is to improve the current solution, introducing an intelligent system based on Machine Learning, which, starting from the classification of events, is able to identify in the text the factors that lead to the occurrence of an accident. In particular, the identification is based on the Semantic Text Similarity between the events described and the events previously collected and analysed. A prototype of an Artificial Neural Network has been developed for this purpose and it has been trained to build three different models, used to evaluate which paradigm is the most effective in this particular context. Semantic Similarity Measures are an important part in Natural Language Processing tasks. However, sentences similarity models are still an immature research field. This thesis introduces innovative elements to standard paradigms, trying to improve the performances of the implemented models in terms of accuracy. Among all, the improving factors are: • The integration with a domain-specific context. • The enhanced Lemmatization of text in the pre-processing phase. • Leveraging available existing data that speed up the factors’ identification process.