Abstract: With the rapid development of large-scale language models,how to reduce the number of model parameters while ensuring model performance has become an important challenge in the field of natural language processing.However,the existing parameter compression techniques are often difficult to balance the stability and generalization ability of the model.To this end,this paper proposes a new framework for sentiment analysis that integrates topic features,aiming to use topic information to enhance the model’s ability to judge text sentiment polarity.Specifically,a method combining LDA and K-means is used to extract the topic features of the text,and it is spliced with word embeddings as a fixed-dimensional vector to obtain a new word vector representation.Sentence-level representation vectors are then constructed using average pooling techniques and fed into a fully connected layer for sentiment classification.To verify the effectiveness of the proposed model,comparative experiments with multiple benchmark algorithms are carried out on public sentiment analysis datasets.Experimental results show that the proposed model is significantly better than ALBERT in multiple data sets,with an accuracy rate increases by about 3.5%,and it maintains high stability and generalization ability with only a small increase in the number of parameters.

Key words: Emotional analysis, ALBERT model, Latent dirichlet allocation, Theme features, Average pooling