Correlation analysis and text classification of chemical accident cases based on word embedding

Accident precursors can provide valuable clues for risk assessment and risk warning. Trends such as the main characteristics, common causes, and high-frequency types of chemical accidents can provide references for formulating safety-management strategies. However, such information is usually documented in unstructured or semistructured free text related to chemical accident cases, and it can be costly to manually extract the information. Recently, text-mining methods based on deep learning have been shown to be very effective. This study, therefore, developed a text-mining method for chemical accident cases based on word embedding and deep learning. First, the word2vec model was used to obtain word vectors from a text corpus of chemical accident cases. Then, a bidirectional long short-term memory (LSTM) model with an attention mechanism was constructed to classify the types and causes of Chinese chemical accident cases. The case studies revealed the following results: 1) Common trends in chemical accidents (e.g., characteristics, causes, high-frequency types) could be obtained through correlation analysis based on word embedding; 2) The developed text-classification model could classify different types of accidents as fires, explosions, poisoning, and others, and the average p (73.1%) and r (72.5%) of the model achieved ideal performance for Chinese text classification; 3) The developed text-classification model could classify the causes of accidents as personal unsafe act, personal habitual behavior, unsafe conditions of equipment or materials and vulnerabilities management strategy; p and r were 63.6% for the causes of vulnerabilities management strategy, and the average p and r are both 60.7%; 4) the accident precursors of explosion, fire, and poisoning were obtained through correlation analyses of each high-frequency type of chemical accident case based on text classification; 5) the text-mining method can provide site managers with an efficient tool for extracting useful insights from chemical accident cases based on word embedding and deep learning.