Search

Pre-training in the cleaned Chinese conversation dataset have problems of reduced generalization ability of the conversation model and low evaluation index after fine-tuning. The method of confrontation training is used to propose a Chinese GPT dialogue model incorporating adversarial training. The projection gradient descent method is used for training during the fine-tuning process, and then the Focal loss function is used to speed up the training. The experimental results show that after integrating confrontation training, fine-tuning and testing in the noisy data set, the model reduces the interference factors such as noise. The experimental results show that after the fusion of adversarial training, the model has improved performance compared with the baseline model, and has strong anti-interference and generalization capabilities. [ABSTRACT FROM AUTHOR]

Aiming at the sparseness and low prediction accuracy of traditional collaborative filtering algorithms, this paper proposed a hybrid recommendation algorithm based on restricted Boltzmann machine and weighted Slope One. Firstly, it used the preliminary filling of the scoring matrix by the restricted Boltzmann machine to alleviate the sparseness problem of the data. Then, it introduced the project attribute information through a hybrid project similarity calculation method. Finally, it adopted the second prediction by the weighted Slope One algorithm to improve the recommended effect. Experiments on the MovieLenslOOK dataset show that the combination of the two algorithms increases the accuracy of the recommendation. [ABSTRACT FROM AUTHOR]

In a class-imbalanced dataset, since the samples close to the class boundary are more likely to be misclassified, it is of great significance to accurately identify boundary samples for classification. Existing methods usually use K-nearest neighbors to identify boundary samples, but the accuracy needs to be improved. To address the above problem, this paper proposed a synthetic oversampling method for boundary minority samples based on Tomek links. This method first found inter-class samples with the nearest distance to form Tomek links. Then, it identifies those minority samples located at the inter-class boundary according to Tomek links. Next, it used the linear interpolation mechanism in synthetic minority oversampling technology (SMOTE) to perform oversampling between the boundary samples and their minority neighbors, thereby achieving the balance of the datasets. The comparison experiment with eight sampling algorithms shows that the proposed method can obtain higher G-mean and F, values on most of the datasets. [ABSTRACT FROM AUTHOR]

Synthetic minority oversampling technology（SMOTE） is one of the effective methods to solve the classimbalanced problem. However, the linear interpolation mechanism of SMOTE restricts the synthesized samples to the connecting line of the original samples, resulting in a lack of diversity for new samples, and may generate noisy samples when this line passes through the majority class region. In response to the above issues, this paper proposed a generation mechanism for minority samples with hypercuboid constraints. This mechanism constructed a hypercuboid as the generation region of new samples instead of linear interpolation, thereby increasing the variability between the synthesized samples and the original samples. Then, it detected whether there were majority samples in the hypercuboid to determine whether to adjust the hypercuboid, which aimed at preventing the new samples into the region of the majority class. This paper integrated the proposed mechanism into three oversampling methods, i. e., SMOTE, Borderline-SMOTE, and ADASYN, by using it to replace linear interpolation, and then experimentally evaluated the integrated methods on 11 benchmark datasets from KEEL. The results showed that compared to the original methods, the integrated methods could help the classifier to obtain higher F1 and comparable G-mean. It verifies that the hypercuboid generation mechanism can significantly improve the classifier’s ability to recognize minority samples, and meanwhile the majority samples are also taken into account. [ABSTRACT FROM AUTHOR]

The BiLSTM neural network model has long training time and cannot fully learn text sentiment information. In order to solve the problems, this paper proposed a text sentiment classification model based on BiGRU-attention neural network. Firstly, it used the bidirectional gated recurrent unity (BiGRU) neural network layer to extract the features of the deep text information. Secondly, it used the attention mechanism (attention) layer to allocate the corresponding weight of the extracted text deep information. Finally, it put the text feature messages of different weights into the softmax function layer to carry out the text sentiment classifications. The experimental results on the IMDB data set show that the accuracy of the proposed neural network model is 90.54%, the loss rate is 0. 2430 and the time cost is 1 100 s and the validity of the BiGRU-attention model is verified. [ABSTRACT FROM AUTHOR]