Semi-supervised Learning Based Entity Recognition Method for Power Dispatch Operation Statements

With the continuous development of the power grid, power grid dispatching faces the problems of complicated business data, inefficient information retrieval and time-consuming operation ticket preparation. The existing operation ticket generation method has been difficult to meet the safe operation of the power grid, and the time-consuming retrieval of the paper dispatch manual has led to the difficulty of efficient retrieval of the dispatch response programme and the inefficiency of the operation ticket preparation. In order to shorten the time consuming operation ticket generation and improve the efficiency of dispatchers, it is necessary to accurately identify the entities of key operation statements such as power dispatch operation steps, equipment, operation time, etc. in the operation ticket document to improve the retrieval efficiency. Therefore, this paper proposes a deep entity recognition model based on scheduling operation statements, which firstly employs a pre-training model, ERNIE, to capture entity-level semantic information, and secondly explores the entity recognition knowledge in the unlabelled data through the proposed semi-supervised training architecture and the loss function scheduling mechanism, so as to improve the performance of the power scheduling entity recognition. In order to verify the effectiveness of this approach, experiments are conducted in this paper using the power dispatch operation text dataset, and the experimental results show that the proposed semi-supervised power dispatch entity recognition method improves the entity recognition precision, recall and F1 score of power dispatch operation statements by 4.43%, 2.06% and 3.24%, respectively, under the full amount of annotated data, as compared to the previous optimal baseline based on fully supervised learning. 3.24%. In particular, with only 10% of labelled data, the F1 score is improved by 15.92% compared to the optimal baseline. These results fully demonstrate the effectiveness of mining knowledge from unlabelled text for power dispatch entity recognition.