A Cascade Binary Tagging Joint Extraction Method

Extracting entities and relations from unstructured text has become a significant task in natural processing, especially knowledge graphs. However, the traditional relationship extraction method processes this task in a pipelined manner, i.e., identifying entity first and then recognizing their relations, and rarely consider the relevance and independence of the two subtasks. To solve the problems of error accumulation, we present a jointly learning method to a tagging problem to tackle this problem. Meanwhile, to understand the semantics of sentences, we introduce the pre-training method into our models, such as BERT. To solve the problem of entity overlap, we replace softmax with a sigmoid activation function. In the language and intelligent technology competition, experimental results show that our model outperforms the baseline model.