Joint Extraction model of Chinese Recipe Entities and Relations Based on Adversarial Training and Cascaded Binary Tagging

The process of creating a knowledge base or knowledge graph relies heavily on the extraction of relational triples from unstructured text. However, the existing methods rarely address the Chinese multi-triple and overlapping triples. We adopt a cascade binary Tagging framework combined with adversarial training (BertAdvCasLSTM) to solve this problem. When the subject information and sentence features are fused, we use a Bi-LSTM network to further feature extraction of the fused sentence features, which improves the performance of the model. We also personally crafted the Recipe_Chinese dataset to ensure the model's extraction capabilities. On both the Recipe_Chinese and DuIE1.0 Chinese datasets, our suggested BertAdvCasLSTM model outperforms the baselines. In-depth experimental analysis is carried out on different cases where a piece of data contains multiple triples, and the performance is improved in most cases. The impact of adversarial training on the overall model performance is also examined.