An extractive reading comprehension model based on a pre-training model with bidirectional attention flow is constructed for the problem of low accuracy in answer prediction that occurs in extractive reading comprehension tasks of service robots. The model first uses a pre-training model to extract the shallow joint semantic representations of the question and the document context, then uses a bidirectional attention network to enhance feature interaction and information fusion to obtain the deep joint semantic features of the question and the document context. Finally, combines the shallow and deeps joint semantic representations to complete the extraction of answers through ranking, error filtering and localization operations. Experiments were conducted on the Stanford English machine reading comprehension dataset SQuAD 1.1 and the "iFlytek Cup" Chinese machine reading comprehension dataset CMRC 2018 for the extractive question and answer task. The results show that compared with the English pre-trained language model BERT, the performance metrics E_M and F1 values of this model are improved by 1.172% and 1.194%, respectively; compared with the Chinese pretrained language model RoBERTa-wwm-ext, the E_M and F1 values are improved by 1.336% and 0.921%, respectively.