Fast Model for Joint Extraction of Entity and Relation

Extracting entities and relations from plain text is a key technique for knowledge and question answering tasks. The traditional multi-head model predicts the relation type of all segment pairs, while the number of negative labels for segment pairs is much larger than positive labels due to the sparsity of relations. At the same time, this calculation method causes the calculation amount to be proportional to the square of sentence length, which reduces the practicability of the model. To solve this problem, a fast extraction model of entity and relation is proposed. For the named entity recognition task, the start and end labels of entities are predicted by two pointer networks, respectively. Semantic segment pairs that do not contain entity end tags are removed in the relation extraction task. This method reduces the number of segment pairs and speeds up inference for relation extraction tasks. To demonstrate the effectiveness of the model, experiments are conducted on the English news dataset ACE05 and the Dutch real estate dataset DREC. The experimental results show that the model achieves competitive performance compared with the baseline model, and its inference speed is improved by 1.4 times on ACE05 and 2.1 times on DREC. © 2023 Journal of Computer Engineering and Applications Beijing Co., Ltd.; Science Press. All rights reserved.