Aspect sentiment triplet extraction (ASTE) is a rapidly emerging task in the field of NLP. Its purpose is to extract aspects and their related opinion expressions and sentiment polarities from review sentences. Although span-level methods, by enumerating all possible spans in the triplet extraction task (ASTE), have achieved good results, how to extract more effective interlingual information and fully utilize interlayer information remains an unresolved issue. Therefore, we propose a Hybrid of Spans and Dual-encoder (HSADE) model that combines spans with dual-encoder. Specifically, we propose a model mainly composed of BERT base encoder and a special encoder made up of a BiLSTM network and a Graph Neural Network (GCN). The base encoder primarily obtains the basic semantics of the language, while the special encoder primarily extracts deeper lexical information as well as syntactic information. An interactive fusion layer is used to integrate the semantics of the base and special encoders, fully utilizing the information between layers. Then, the fused information obtained is used to compute the span representation through a span formula. In decoding, we defined eight types of labels to identify aspects, opinions, and sentiments. Finally, we carried out extensive tests on four standard datasets, and the outcomes of these experiments validated the effectiveness of the HSADE model.