IEEE Access (Jan 2024)
ADHN: Sentiment Analysis of Reviews for MOOCs of Dilated Convolution Neural Network and Hierarchical Attention Network Based on ALBERT
Abstract
Massive Open Online courses (MOOCs) are increasingly utilized by learners for knowledge acquisition and skill development. Accurate extraction of emotional information from MOOC course reviews plays a pivotal role in enhancing the quality of MOOC courses and fostering sustainable growth of MOOC platforms. Currently, sentiment analysis of MOOC course reviews predominantly focuses on general aspects, overlooking the hierarchical structure of text. Moreover, recurrent neural networks suffer from recursion limitations leading to reduced computational efficiency, while word embedding fails to address the issue of one-time polysemy. In this study, we propose ALBERT-DCNN-HAN (ADHN), an advanced text sentiment analysis model based on Dilated Convolution Neural Network and Hierarchical Attention network derived from ALBERT (A Lite BER) Network. The model primarily relies on the continuous updating of DCNN to compensate for the lack of hierarchical structure in deep neural networks, while also addressing the conflict between traditional word segmentation techniques and the trend towards emotion expression. Firstly, we employ the ALBERT model to generate ALBERT word vectors that integrate contextual features and dynamic semantics. ALBERT further incorporates contextual features from the sentence in which each word is located into its corresponding word vector, thereby generating distinct semantic vectors based on different meanings of polysemous words. Subsequently, these ALBERT word vectors are sampled and computed using DCNN to extract text features across multiple scales of context. Moreover, in order to capture both sentence-level and word-level characteristics comprehensively during text emotion expression, we fully consider hierarchy by integrating a hierarchical attention mechanism. Finally, Conditional random field (CRF) is employed for emotion prediction. Through analysis of information from empirical dataset derived from reviews of MOOC courses, our results demonstrate that this model effectively extracts valuable textual information and achieves an improved bias classification accuracy on review datasets compared with other neural network models.
Keywords
