Construction is an industry well known for its very high rate of injuries and accidents around the world. Even though many researchers are engaged in analysing the risks of this industry using various techniques, construction accidents still require much attention in safety science. According to existing literature, it has been found that hazards related to workers, technology, natural factors, surrounding activities and organisational factors are primary causes of accidents. Yet, there has been limited research aimed to ascertain the extent of these hazards based on the actual reported accidents. Therefore, the study presented in this paper was conducted with the purpose of devising an approach to extract sources of hazards from publicly available injury reports by using Text Mining (TM) and Natural Language Processing (NLP) techniques. This paper presents a methodology to develop a rule-based extraction tool by providing full details of lexicon building, devising extraction rules and the iterative process of testing and validation. In addition, the developed rule-based classifier was compared with, and found to outperform, the existing statistical classifiers such as Support Vector Machine (SVM), Kernel SVM, K-nearest neighbours, Naïve Bayesian classifier and Random Forest classifier. The finding using the developed tool identified the worker factor as the highest contributor to construction site accidents followed by technological factor, surrounding activities, organisational factor, and natural factor (1%). The developed tool could be used to quickly extract the sources of hazards by converting largely available unstructured digital accident data to structured attributes allowing better data-driven safety management.