English Composition Topic and Emotion Classification Based on LDA-BERT Fusion and Dual GRU model

Abstract

With the development of globalization and artificial intelligence, the traditional manual correction of English compositions has problems, such as low efficiency and a lack of standardized evaluation. Therefore, there is an urgent need for automated analysis. The study aims to develop a precise model for identifying English essay themes and classifying emotional perspectives. Latent Dirichlet allocation is used to combine Probase prior knowledge and BERT fusion technology for topic modeling, and a Bi-GRU network with an attention mechanism is used for emotion classification. The experiment was carried out based on a dataset of 66,000 English compositions, including TOEFL, IELTS, and classroom compositions, and verified the effectiveness and superiority of the joint model in topic identification and emotion classification tasks. The experiment showed that the topic recognition model had a perplexity of 803, coherence of 0.831, and inter-class distance of 3.2 when there were 25 topics. When the threshold of the emotion classification model was 0.45, the accuracy reached 80.1% and the F1 value was 73.7%. The F1 value of 2,000 tests was 76.1%, which was better than the comparison model. Research has shown that this solution effectively solves problems such as insufficient semantic understanding in existing technologies, provides scientific support for intelligent evaluation systems, and promotes the intelligent development of educational evaluation.