Efficient Relation Extraction of Automobile Faults Based on Hidden Knowledge in Prompt Tuning Templates

In this paper, we propose a relation extraction method based on the Prompt Tuning (PT) technique, which aims to improve the accuracy and efficiency of knowledge extraction in automotive fault texts. The method can be effectively applied to long-tail or less resourceful situations by mining the hidden knowledge in templates. Experimental results demonstrate that the method performs better in dealing with long-tailed data with a minimal small of relation labels, successfully overcoming issues such as catastrophic forgetting and misrecognition. In a data ratio of approximately 15:1, the predicted F1 values for relation labels (representing performance failures) are improved by 4.15%, 1.24%, 1.26%, and 0.03%, respectively, compared to the best results obtained from the comparison model. Finally, the extracted triples (subject, relation, object) are imported into the Neo4j graph database to realize intelligent retrieval of automotive faults.