Sentiment Classification Based on Neural Networks and Data Augmentation

Sentiment classification is a basic and critical task in natural language processing, which is widely used in the fields of public opinion analysis and product evaluation. Traditional sentiment classification methods relying on manual annotation are not only inefficient and costly, but also difficult to adapt to the needs of large-scale text processing. In recent years, pre-trained language models based on BERT have made significant progress in sentiment classification, but there is still room for improving their classification accuracy when dealing with fine-grained semantic scenarios such as Chinese hotel reviews. To address these challenges, I propose a sentiment classification method that combines data enhancement with neural network structure improvement. Specifically, BiLSTM and MLP modules are introduced based on BERT representations to enhance contextual modeling capabilities, and a static rule-based or context-aware dynamic data augmentation strategy is applied to augment textual data, aiming to improve the generalization ability of the model. An empirical study conducted on the publicly available Chinese sentiment analysis corpus ChnSentiCorp dataset [1] demonstrates that the proposed method effectively enhances classification performance, improving the accuracy from 88.17% (baseline) to 90.00%. The experimental results validate the effectiveness and robustness of the mentioned data enhancement and model structure modification approach in the Chinese sentiment analysis task.