Overlapping relational triple extraction refers to identifying relational triples within a sentence that share entities with each other. Due to its more challenging and increasing realistic, it has become the focus of knowledge extraction at present. Most previous works on extracting overlapping relational triples have used tagging based methods to identify pairs of entities and subsequently predict their relations. However, the effect of such methods is severely affected by the incompleteness and inaccuracy of entity pair recognition. To overcome these problems, we introduce a joint model based global detection and bidirectional tagging (GDBT). Our model significantly enhances the completeness of entity pair recognition based on bidirectional tagging and further improves the accuracy of entity pair through global detection. In the relation extraction stage, an entity attention network is elaborated to infer relations between entities, and negative sampling strategy is adopted to augment the model's capacity to differentiate between negative entity pairs, thus enhancing the model's generalization. Extensive experiments on public datasets have shown that GDBT achieves significant performance improvement and efficiently handles the challenge of overlapping relational triples compared to prevailing methods.