ACM Transactions on Asian and Low-Resource Language Information Processing最新文献

In the field of Natural Language Processing (NLP), Deep Learning (DL) and Neural Network (NN) technologies have been widely applied to machine translation and sentiment analysis and have demonstrated outstanding performance. In recent years, NLP applications have also combined multimodal data, such as visual and audio, continuously improving language processing performance. At the same time, the size of Neural Network models is increasing, and many models cannot be deployed on devices with limited computing resources. Deploying models on cloud platforms has become a trend. However, deploying models in the cloud introduces new privacy risks for endpoint data, despite overcoming computational limitations. Federated Learning (FL) methods protect local data by keeping the data on the client side and only sending local updates to the central server. However, the FL architecture still has problems, such as vulnerability to adversarial attacks and non-IID data distribution. In this work, we propose a Federated Learning aggregation method called FedREAS. The server uses a benchmark dataset to train a global model and obtains benchmark updates in this method. Before aggregating local updates, the server adjusts the local updates using the benchmark updates and then returns the adjusted benchmark updates. Then, based on the similarity between the adjusted local updates and the adjusted benchmark updates, the server aggregates these local updates to obtain a more robust update. This method also improves the client selection process. FedREAS selects suitable clients for training at the beginning of each round based on specific strategies, the similarity of the previous round’s updates, and the submitted data. We conduct experiments on different datasets and compare FedREAS with other Federated Learning methods. The results show that FedREAS outperforms other methods regarding model performance and resistance to attacks.

Knowledge management is crucial to the teaching and learning process in the current era of digitalization. The idea of "learning via working together" is making Natural Language Processing a popular tool to improve the learning process based on the intelligent system for evaluating the composition. English language learning is highly dependent on the composition written by the students under various topics. Teachers are facing huge difficulties in the evaluation of the composition as the level of writing by the students will vary for individual. In this research, Natural Language Processing concept is utilized for getting trained with the student's writing skills and Multiprocessor Learning Algorithm (MLA) combined with Convolutional Neural Network (CNN) (MLA-CNN) for evaluating the composition and declaring the scores for the students. The model's composition scoring rate is validated using a range of learning rate settings. Some theoretical notions for smart teaching are proposed, and it is hoped that this automatic composition scoring model would be used to grade student writing in English classes. When applied to the automatic scoring of students' English composition in schools, the suggested composition scoring system trained by the MLP-CNN has great performance and lays the groundwork for the educational applications of ML inside AI. The study results proved that the proposed model has provided an accuracy of 98%.