基于TRANSFORMER的文本情感分类方法研究

Emotion classification is a classification technology with great practical value. It is widely used in some real scenes, such as film box office prediction, and has attracted much attention all the time. In order to explore the performance and defects of the current mainstream deep learning methods in text emotion classification tasks, this paper compares and evaluates several mainstream methods based on transformer, including Bert and its improved models: Roberta, distilbert and minilm. After the experiment on IMDB film review emotion classification task, it is found that the current multilingual pre training method will reduce the classification performance of Bert model; Different from the visual model and the language model simplified by distillation, its classification ability will decline slightly; Roberta's training method is excellent and worthy of in-depth study. This paper provides a further improvement direction for Bert emotion classification.

情感分类（Emotion Classification）是一项极具实用价值的分类技术，广泛应用于电影票房预测等诸多真实场景中，长期以来备受关注。为探究当前主流深度学习方法在文本情感分类任务中的性能与缺陷，本文对多款基于Transformer的主流方法展开对比与评估，其中包括BERT及其改进模型：RoBERTa、DistilBERT与MiniLM。在IMDB电影评论情感分类任务上开展实验后，本文发现：当前的多语言预训练方法会降低BERT模型的分类性能；与视觉模型及经蒸馏简化的语言模型不同，DistilBERT的分类能力仅出现小幅下降；RoBERTa的训练方法表现优异，值得开展深入研究。本文为BERT情感分类任务提供了进一步的优化方向。