Construction of an English Essay Topic and Opinion Analysis Model Based on PLDA-NTM and Bi-GRU

As a global lingua franca, English plays a crucial role in expanding Chinese students’ knowledge and broadening their horizons. English writing, as an important output method of language competence, has attracted widespread attention. However, the “one-to-many” model in current teaching limits teachers’ timely and accurate guidance on each essay, hindering the improvement of students’ writing skills. With the development of artificial intelligence technology, automatic essay scoring systems have emerged, which can alleviate teachers’ burdens while improving feedback efficiency. Nevertheless, existing systems are insufficient in topic mining and sentiment analysis, restricting the accuracy of content evaluation. Therefore, constructing a model capable of in-depth analysis of essay topics and sentiments holds significant practical value. This paper proposes an English essay analysis model integrating topic semantics and sentiment opinions. By extracting deep-level topic information of essays, combining topic-based sentiment discrimination, and calculating semantic similarity at the sentence level, the model achieves a comprehensive evaluation of essay content quality. The specific research contents include: (1) Constructing a contextual topic recognition model integrating conceptual knowledge (PLDA-NTM), which uses the Probase concept base to enhance the semantic representation ability of the LDA model and combines BERT to generate local semantics, thereby improving topic clustering quality. (2) Establishing a topic-based sentiment opinion analysis model (TA-GRU), which adopts a Bi-GRU network to extract contextual features and identifies sentiment tendency categories through a topic-level attention mechanism. (3) Designing a topic and opinion quality analysis module that calculates topic similarity, content relevance, and sentiment quality at the sentence level, and generates scores and comments through weighting. (4) Verified by multiple sets of corpus experiments, the proposed model demonstrates higher accuracy and practical value in automatic scoring.