Multi-BERT-wwm Model Based on Probabilistic Graph Strategy for Relation Extraction
| Autor: | Yunpeng Cai, Yingxiang Zhang, Qiong Wang, Li Chen, Xiangyun Liao, Hongjun Kang |
|---|---|
| Rok vydání: | 2021 |
| Předmět: | |
| Zdroj: | Health Information Science ISBN: 9783030908843 HIS |
| Popis: | As the core work of information extraction, relation extraction aims to find medical entity pairs with relations from the medical records. The current methods of relation extraction either ignore the relevance of entity extraction and relation classification, or fail to solve the problem of multiple relations or entities in one sentence. To handle those problems, this paper proposes a cascading pointer Multi-BERT-wwm model based on the probabilistic graph strategy. The model selects an entity randomly from all the predicted entities each time, then predicts other entities and their relations according to that entity. Meanwhile, the pointer labeling network helps to solve the problem of overlapping entities. The Multi-BERT-wwm model is improved based on BERT, which connects a layer of self-attention to the multi-head attention layer in the first six Encoder modules to strengthen the feature extraction ability. In addition, we add the adversarial training to improve the robustness and generalization ability of the model. The experimental results tested on the CMeIE dataset show that compared with the traditional CNN+Attention and BERT methods, our method improves the F1-score by 4.42% and 1.91% respectively in relation extraction task. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|