The development of pre-trained language models has greatly promoted the progress of machine reading comprehension tasks. In order to make full use of shallow features of the pre-trained language model and further improve the accuracy of predictive answer of question answering model， a three-stage question answering model based on Bidirectional Encoder Representation from Transformers （BERT） was proposed. Firstly， the three stages of pre-answering， re-answering and answer-adjusting were designed based on BERT. Secondly， the inputs of embedding layer of BERT were treated as shallow features to pre-generate an answer in pre-answering stage. Then， the deep features fully encoded by BERT were used to re-generate another answer in re-answering stage. Finally， the final prediction result was generated by combining the previous two answers in answer-adjusting stage. Experimental results on English dataset Stanford Question Answering Dataset 2.0 （SQuAD2.0） and Chinese dataset Chinese Machine Reading Comprehension 2018 （CMRC2018） of span-extraction question answering task show that the Exact Match （EM） and F1 score （F1） of the proposed model are improved by the average of 1 to 3 percentage points compared with those of the similar baseline models， and the model has the extracted answer fragments more accurate. By combining shallow features of BERT with deep features， this three-stage model extends the abstract representation ability of BERT， and explores the application of shallow features of BERT in question answering models， and has the characteristics of simple structure， accurate prediction， and fast speed of training and inference.

Concerning that the traditional saliency detection algorithm has low segmentation accuracy and the deep learning-based saliency detection algorithm has strong dependence on pixel-level manual annotation data， an unsupervised saliency object detection algorithm based on graph cut refinement and differentiable clustering was proposed. In the algorithm， the idea of “coarse” to “fine” was adopted to achieve accurate salient object detection by only using the characteristics of a single image. Firstly， Frequency-tuned algorithm was used to obtain the salient coarse image according to the color and brightness of the image itself. Then， the candidate regions of the salient object were obtained by binarization according to the image’s statistical characteristics and combination of the central priority hypothesis. After that， GrabCut algorithm based on single image for graph cut was used for segmenting the salient object finely. Finally， in order to overcome the difficulty of imprecise detection when the background was very similar to the object， the unsupervised differentiable clustering algorithm with good boundary segmentation effect was introduced to further optimize the saliency map. Experimental results show that compared with the existing seven algorithms， the optimized saliency map obtained by the proposed algorithm is closer to the ground truth， achieving an Mean Absolute Error （MAE） of 14.3% and 23.4% on ECSSD and SOD datasets， respectively.

With the development of social networks， it has important social value to analyze the sentiments of massive texts in the social networks. Different from ordinary text classification， short text sentiment classification needs to mine the implicit sentiment semantic features， so it is very difficult and challenging. In order to obtain short text sentiment semantic features at a higher level， a new Multi-head Attention Memory Network （MAMN） was proposed for sentiment classification of short texts. Firstly， n-gram feature information and Ordered Neurons Long Short-Term Memory （ON-LSTM） network were used to improve the multi-head self-attention mechanism to fully extract the internal relationship of the text context， so that the model was able obtain richer text feature information. Secondly， multi-head attention mechanism was adopted to optimize the multi-hop memory network structure， so as to expand the depth of the model and mine higher level contextual internal semantic relations at the same time. A large number of experiments were carried out on Movie Review dataset （MR）， Stanford Sentiment Treebank （SST）-1 and SST-2 datasets. The experimental results show that compared with the baseline models based on Recurrent Neural Network （RNN） and Convolutional Neural Network （CNN） structure and some latest works， the proposed MAMN achieves the better classification results， and the importance of multi-hop structure in performance improvement is verified.