نوع مقاله : مقاله پژوهشی
نویسندگان
گروه کامپیوتر و نرمافزار، دانشکده مهندسی کامپیوتر، دانشگاه یزد، یزد، ایران.
چکیده
کلیدواژهها
موضوعات
عنوان مقاله [English]
نویسندگان [English]
Aspect-based sentiment analysis of tweets with deep neural networks is of great interest. However, these networks require rich data for optimal performance. Therefore, this study uses zero-shot deep transfer learning and cross-domain data to increase the richness of educational resources in the target domain. First, we collected the tweets of cryptocurrency influencers and then used the combination of the deep neural network, attention layer, and pre-trained Distillbert for aspect-based sentiment analysis. The heat map of the attention layer shows that using this layer after deep models has helped highlight the aspect words. In zero-shot transfer learning, the model does not have samples of the labelled target data, and training only uses cross-domain data. In addition, using cross-domain data and their optimal selection with Pearson's similarity coefficient in conditions of lack of rich data reduces the negative transfer. It helps the model's sensitivity to the context of the text. Also, the Synthetic Minority Over-sampling Technique has been used to solve the problem of unbalanced data. The experimental findings show that the model has increased the average accuracy and F1 on the SemEval benchmark dataset by about 2%, and the average ROC-AUC rate has reached 86.35% compared to previous research. Also, we used the adaptive learning rate to accelerate the model's convergence and Grid Search to select the network's hyperparameters, making the model more adaptable to the target domain.
کلیدواژهها [English]
)