KeySim-FLMacBERT: A Law Recommendation Algorithm for Hard Example Mining in Work Safety

There is a complex legal system in work safety, but currently only manual means are used to recommend law, and intelligent methods are urgently needed. To address law recommendation for hard examples, we propose a law recommendation model using data augmentation named KeySim-FLMacBERT, which is an improved version of MacBERT and is optimized from both data and algorithm perspectives: KeySim module leverages a keyword library of work safety laws to enhance the illegal fact, and uses optimized SimBERT text generation method to achieve hard example mining; FLMacBERT module adopts a priori guidance-based weight optimization strategy to optimize α in Focal Loss instead of cross entropy. Experiments prove the feasibility of our model on a real dataset of illegal facts and law of work safety in Beijing. The weighted F1 score of our model on all samples is 96.64%, and the F1 score on hard examples is 99.22%, exceeding the cross entropy by 3.06%.