A Simple Completely Adjacency List Oriented Relational Extraction Model

Abstract

Entity relationship extraction aims to extract important triplet information from massive unstructured data, which is the basis of downstream tasks such as building a knowledge map. The graph perspective is used to analyze the entity and relationship extraction and build adjacency list oriented model to solve the problem of large space consumption of the adjacency matrix, but it uses complex operations to extract entities and relationships sequentially. Therefore, we propose a simple completely adjacency list oriented relationship extraction model. This model firstly introduce a realtion label-aware module to supplement sentence information and a feature separation module to alleviate the error accumulation problem caused by sequential extraction, and then sequentially extracts subjects, objects, and relations. Extensive experiments on two common datasets have shown that our model maintains high accuracy of 92.8while also significantly improving inference speed down from 35.7ms to 22.9ms.