A Study of Large Model Sentiment Classification Based on LoRA Fine-Tuning

Sentiment classification is an important task in natural language processing, which is widely used in many fields such as public opinion monitoring, commodity evaluation, and social media analysis. In recent years, pre-trained large language models have performed prominently in the task of sentiment categorization, but large language models still have certain deficiencies in dealing with the complexity and polysemy of textual sentiments as well as long textual reasoning. With the emergence of deepseek model, this paper, based on the Chinese microblog sentiment classification dataset, adopts LoRA (Low-Rank Adaptation) fine-tuning technique with the help of deepseek model to improve its performance of sentiment classification, and compares it with the effect of the model before fine-tuning. The experimental results show that the LoRA fine-tuning method exhibits excellent results in improving the model in sentiment classification tasks, especially in processing complex text and binary sentiment labeling tasks with stronger robustness. The finetuned model improves 14.10%, 12.82% and 12.85% in Accuracy, F1-Score and ROC-AUC metrics, respectively, compared with the baseline model, which effectively improves the accuracy and diversity of complex text sentiment classification.