Abstract: Aimed at the problems of entity overlap, error accumulation and lack of dependency in traditional information extraction methods, a joint model of entity recognition and relationship extraction based on fragment arrangement and multi-head selection is proposed. We established the dependence between entity recognition and relationship extraction by sharing the coding layer, and solved the problem of entity overlap at the level of span through fragment arrangement. The multi-head selection mechanism was used to predict the relationship between entities, and the confrontation training was added, which was constrained by the auxiliary loss function. Through ablation experiments and experiments based on different weight loss functions, the best parameters were found. The model was implemented in the Chinese dataset DulE 2.0, and the F1 value of 0.829 was achieved, which was increased by 2.24% compared with the baseline model with the best effect.