Enity Relation Extraction of Industrial Robot PHM Based on BiLSTM-CRF and Multi-head Selection.

This paper extracted the entity relation from unstructured text in the industrial robot prognostics and health management (PHM) data for the construction of a knowledge graph. Traditionally, there is a disadvantage of error propagation in the pipeline method. At the same time, the industrial robot PHM corpus has a relation overlap problem, which reduces the accuracy of relations extraction. To solve this problem, we proposed a joint entity relations extraction model combining BiLSTM-CRF and multi-head selection. In this model, the encoding layer is shared by entity extraction tasks and relations extraction tasks. Moreover, BiLSTM-CRF is introduced for named entity recognition tasks, and the entity recognition information is utilized by the multi-head selection structure to solve the problem of overlapping relations. The results show that the model can effectively extract the entity relations of unstructured text in the robot PHM data, and the overall F1-score for entity relations extraction reaches 87.64. It is an increase of 2.92% and 13.5%, compared with the BiLSTM-ED-CNN model and the pipeline method using the same model.