Abstract: Question Answering is a challenging task in Natural Language Processing, ‘which can be classified into single-hop Question Answering and multi-hop Question Answering. To address the more challenging multi-hop question answeringtask, this paper proposes a method based on pretrained language models for the HotpotQA explainable multi-hop question answering dataset. This method applies the classical Divide and Conquer strategy to split the multi-hop questions into simpler single-hop sub questions. These sub questions are encoded using a pretrained language model with unsupervised contrastive learning, and data augmentation is applied to improve the question decomposition accuracy. In the process of answer generation, we proposes a window self-attention mechanism and adopt a “two-step filtering” method to retrieve the supporting facts and the answers. Experiments on HotpotQA distractor dataset demonstrate that the proposed method outperformed other 12 alternatives by achieving 48.10% EM score.