A Chinese Entity-Relation Extraction Method Via Improved Machine Reading Comprehension

As the downstream task of building a knowledge graph, Chinese entity relationship extraction from unstructured texts plays an important role in the field of natural language processing. There are two main ways for Chinese entity relation extraction: joint extraction method and pipeline extraction method. The joint extraction method outputs the relation triples contained in the texts directly in a row, which causes two problems: the lack of external knowledge and the nesting of entities. This article proposes a method to take the advantage of the similarity between the span extraction task and the information extraction task, and transforms entity relation extraction problem into a task similar to machine reading comprehension. This method first uses the Roberta pre-training model to obtain word representation with relation information, and then identifies entity pairs that may exists under per relation through a global pointer network, which outperforms better than the normal pointer network. By comparing different models’ performance on the same dataset, the results show the accuracy, recall and F1 scores of our method are higher than other methods, which proves the effectiveness of our method.