The neural network models based on attention mechanism are mainly used in the field of aspect-level sentiment analysis. The dependencies between aspect words and opinion words， as well as the distances between aspect words and context words， are ignored by this type of models， which further leads to inaccurate classification of emotions by this type of models. To solve above problems， a Relational and Interactive Graph ATtention network （RI-GAT） model was established. Firstly， the semantic features of sentences were learned by the Long Short-Term Memory （LSTM） network. Then the learned semantic features were combined with the position information of sentences to generate new features. Finally the dependencies between various aspects words and opinion words were extracted from the new features， realizing efficient and comprehensive use of syntactic dependency information and position information. Experimental results on Laptop， Restaurant， and Twitter datasets show that compared to the suboptimal Dynamic Multi-channel Graph Convolutional Network （DM-GCN）， RI-GAT model has the classification Accuracy （Acc） improved by 0.67， 1.65， and 1.36 percentage points， indicating that RI-GAT model can better establish the relationship between aspect words and opinion words， making sentiment classification more accurate.

Aiming at the problems of dense knowledge and the propagation of error during entity extraction and relation classification in medical domain， a high-precision entity and relation extraction framework based on pseudo-entity data augmentation was proposed. First， a Transformer-based feature reading unit was added in the entity extraction module to capture category information for accurately identifying medical long entities among dense entities. Second， a relation negative example generation module was inserted into the pipeline extraction framework， pseudo-entities were generated for confusing relation classification model by an under-sampling-based pseudo-entity generation model， and three data augmentation generation strategies were proposed to improve the model’s ability to identify subject-object reversal， subject-object boundary errors， and relation classification errors. Finally， the problem of the sharp increase in training time caused by data enhancement was alleviated by the levitated-marker-based relation classification model. On CMeIE dataset， four mainstream models were compared with the proposed model. For entity extraction tasks， the proposed model improved the F1 value by 2.26% compared with suboptimal model PL-Marker（Packed Levitated Marker）， while for entity relation extraction tasks， the proposed medel improved the F1 value by 5.45% and the precision by 15.62% compared with suboptimal pipeline extraction model proposed by CBLUE （Chinese Biomedical Language Understanding Evaluation）. The experimental results show that using both the feature reading unit and the pseudo-entity data enhancement module can effectively improve the precision of extraction.

To address the difficulty of word boundary recognition stemming from nested entities in Chinese medical texts， as well as significant semantic information loss in existing Lattice-LSTM structures with integrated lexical features， an adaptive lexical information enhancement model for Chinese Medical Named Entity Recognition （MNER） was proposed. First， the BiLSTM （Bi-directional Long-Short Term Memory） network was utilized to encode the contextual information of the character sequence and capture the long-distance dependencies. Next， potential word information of each character was modeled as character-word pairs， and the self-attention mechanism was utilized to realize internal interactions between different words. Finally， a lexicon adapter based on bilinear-attention mechanism was used to integrate lexical information into each character in the text sequence， enhancing semantic information effectively while fully utilizing the rich boundary information of words and suppressing words with low correlation. Experimental results demonstrate that the average F1 value of the proposed model increases by 1.37 to 2.38 percentage points compared to the character-based baseline model， and its performance is further optimized when combined with BERT.

Imbalanced classification is one of the research hotspots in the field of machine learning， where oversampling increases minority samples through repeated extraction or artificial synthesis to rebalance the dataset. However， most of the existing oversampling methods are based on the original data distribution， and are difficult to reveal more dataset distribution characteristics. To address the above problem， firstly， an improved semi-supervised clustering algorithm was proposed to mine the data distribution characteristics. Secondly， based on the results of semi-supervised clustering， the highly-confident unlabeled data （pseudo-labeled samples） was selected from minority-class clusters to join into the original training set. In this way， in addition to rebalancing the dataset， the distribution characteristics obtained by semi-supervised clustering was able to be used to assist the imbalanced classification. Finally， the results of semi-supervised clustering and classification were fused to predict the final labels， which further improved the model performance of imbalanced classification. With G-mean and Area Under Curve （AUC） selected as evaluation indicators， the proposed algorithm was compared with seven oversampling-/undersampling-based imbalanced classification algorithms， such as TU （Trainable Undersampling） and CDSMOTE （Class Decomposition Synthetic Minority Oversampling TEchnique） on 10 public datasets. Experimental results show that compared with TU and CDSMOTE， the proposed algorithm has the average AUC increased by 6.7% and 3.9% respectively， the average G-mean improved by 7.6% and 2.1% respectively. At the same time， the proposed algorithm achieves the highest average results on both evaluation indicators than all the algorithms to be compared. It can be seen that the proposed algorithm can effectively improve the imbalanced classification performance.

Concerning the problem that the linear eigentransformation method cannot capture the statistical properties of the nonlinear facial image, a Data-driven Local Eigentransformation (DLE) method for face hallucination was proposed. Firstly, some samples most similar to the input image patch were searched. Secondly, a patch-based eigentransformation method was used for modeling the relationship between the Low-Resolution (LR) and High-Resolution (HR) training samples. Finally, a post-processing approach refined the hallucinated results. The experimental results show the proposed method has better visual performance as well as 1.81dB promotion over method of locality-constrained representation in objective evaluation criterion for face image especially with noise. This method can effectively hallucinate surveillant facial images.

To solve the Small Sample Size (SSS) problem and the "inferior" problem of traditional Fisher Linear Discriminant Analysis (FLDA) when it is applied to Synthetic Aperture Radar (SAR) image recognition tasks, a new image feature extraction technique was proposed based on weighted two-directional and two-dimensional linear discriminant analysis (W(2D)2LDA). First, the scatter matrices in the two-directional and two-dimensional linear discriminant analysis criterion were modified by adding weights. Then, feature matrix was extracted by W(2D)2LDA. The experimental results with MSTAR dataset verify the effectiveness of the proposed method, and it can strengthen the feature's discrimination and obtain better recognition performance with fewer memory requirements simultaneously.