Research on Domain Text Classification Algorithm Based on Multi-Granularity Fusion

The current research on domain text classification is primarily based on word granularity to facilitate the extraction and representation of text information. However, this approach is susceptible to the influence of out-of-vocabulary words and is insufficient for inter-word and inter-sentence semantic representation. Given the capacity of sentence template to effectively summarize sentence semantics, this paper proposes a domain text classification algorithm based on multi-granularity fusion. The algorithm introduces sentence template information to enhance text features from a semantic perspective, constructs embedding vectors of characters, as well as word and sentence templates. It extracts the contextual semantic features of these vectors using BiLSTM, further identifies key features of the characters and words through the attention mechanism, and finally, the information at each granularity is fused using the weighted fusion module to obtain the final semantic feature vectors. These vectors are then input into the softmax function for classification prediction. Experimental results demonstrate that the sentence template information significantly improves the model's ability to represent text semantics. Additionally, the weighted feature fusion method effectively retains and utilizes more feature information. These two improvements enhance classification accuracy by 1.77% and 1.49%, respectively.