Automatic Cause Inference of Construction Accident Using Long Short-Term Memory Neural Networks

Research of predicting the causes of construction accidents from documents has attracted increased interest in the passing three decades. One main branch of this type of research is to use automatic methods to enable effective causal inference from a large amount of textual data. To improve the accuracy and reduce labor resources required, learning-based methods have been successfully employed over full texts of construction accident reports. However, to date, these methods are not capable of wide application in the construction industry, where most of the accident narratives are recorded as short texts. Moreover, the data imbalance problem is a frequent bottleneck in the classification performance. To achieve a higher degree of adaptability for construction accident classification, this study develops a framework consisting of data augmentation, Bi-LSTM and self-attention neural networks, and focal loss objective function, which is trained and tested over two data sets consisting of short-text and imbalanced data. The validation results showed that, even with much less information provided in the short texts, the proposed model has superior performance to existing methods.