Search

With the informatization of the society, it’s of great practical value to extract useful information from massive textual data available online using tools of NLP (natural language processing). However, the texts collected from social platforms suffer from issues such as low amount of valuable data and data imbalance. This paper proposes two methods to deal with these problems, named SimDyFeFL (SimBERT & dynamic feedback Focal Loss) and EdaDyFeFL（EDA & dynamic feedback Focal Loss）, one is applicable for crisis-related information recognition tasks in Chinese, and another is for cyber trolls detection tasks in English. Specifically, SimBERT and EDA (easy data augmentation) methods are used to augment the original data with large differences between classes to a similar number of classes, and then the Focal Loss function with dynamic feedback process is fused to weight each class. Then, BERT (bidirectional encoder representations from transformers), RoBERTa (robustly optimized BERT pre-training approach), and BERT_DPCNN (BERT deep pyramid convolutional neural networks) text classification models are designed for three-stage comparative experiments to validate the effectiveness of proposed methods. Extensive experiments on two real datasets in Chinese and English show that the performance of the improved text classification models using SimDyFeFL and EdaDyFeFL is significantly improved, the accuracy of Chinese model is increased by 7.70 percentage points, and the accuracy of English model is increased by 5.15 percentage points. Compared with the best results on the Kaggle platform, the accuracy of the English model is 2.92 percentage points higher, and the Macro F1 score and Weighted F1 score are 2.83 percentage points and 2.95 percentage points higher, respectively.

The malicious use of deepfake videos seriously affects information security and brings great harm to society. Currently, deepfake videos are mainly generated based on deep learning methods, which are difficult to be recognized by the naked eye, therefore, it is of great significance to study accurate and efficient deepfake video detection techniques. Most of the existing detection methods focus on analyzing the discriminative information in a specific feature domain for classification from a local or global perspective. Such detection methods based on a single type feature have certain limitations in practical applications. In this paper, we propose a deepfake detection method with the ability to comprehensively analyze the forgery face features, which integrates features in the space domain, noise domain, and frequency domain, and uses the Inception Transformer to learn the mix of global and local information dynamically. We evaluate the proposed method on the DFDC, Celeb-DF, and FaceForensic++ benchmark datasets. Extensive experiments verify the effectiveness and good generalization of the proposed method. Compared with the optimal model, the proposed method with a small number of parameters does not use pre-training, distillation, or assembly, but still achieves competitive performance. The ablation experiments evaluate the role of each component. [ABSTRACT FROM AUTHOR]

With the rise of multi-modal methods, multi-modal knowledge graphs have become a better choice for storing human knowledge. However, knowledge graphs often suffer from the problem of incompleteness due to the infinite and constantly updating nature of knowledge, and thus the task of knowledge graph completion has been proposed. Existing multi-modal knowledge graph completion methods mostly rely on either embedding-based representations or graph neural networks, and there is still room for improvement in terms of interpretability and the ability to handle multi-hop tasks. Therefore, we propose a new method for multi-modal knowledge graph completion. Our method aims to learn multi-level graph structural features to fully explore hidden relationships within the knowledge graph and to improve reasoning accuracy. Specifically, we first use a Transformer architecture to separately learn about data representations for both the image and text modalities. Then, with the help of multimodal gating units, we filter out irrelevant information and perform feature fusion to obtain a unified encoding of knowledge representations. Furthermore, we extract multi-level path features using a width-adjustable sliding window and learn about structural feature information in the knowledge graph using graph convolutional operations. Finally, we use a scoring function to evaluate the probability of the truthfulness of encoded triplets and to complete the prediction task. To demonstrate the effectiveness of the model, we conduct experiments on two publicly available datasets, FB15K-237-IMG and WN18-IMG, and achieve improvements of 1.8 and 0.7%, respectively, in the Hits@1 metric.