A topic joint model for knowledge extraction from unstructured maintenance records

Extracting knowledge and constructing domain knowledge graphs from used vehicle maintenance records can reveal the mechanism of faults. However, maintenance records are semi-structured documents containing both structured and unstructured data. Here, we propose a topic joint model to improve the accuracy of knowledge extraction from used vehicle maintenance records. First, we apply the weighted latent Dirichlet allocation method to extract the hidden topic distribution of fault cases. Then, a bidirectional encoder representations from transformers (BERT) -based method is developed to identify the position and category information of fault entities with case-topic features. Finally, we used real data of vehicle maintenance records to verify the efficacy of our proposed method. The results show that when using real-world data, the proposed joint model has a higher accuracy than the standard knowledge extraction method, which is achieved by combining topic features for entity recognition and relationship extraction tasks. The F1 scores of entity recognition and relationship classification were 75.46 and 78.42, respectively.