Small Sample Classification Based on Prompt Learning

With the development of the information age and the gradual popularization of online shopping, a large amount of transaction information is generated during the transaction process of goods, and the transaction data has enormous analytical value. For the problem of e-commerce text classification, most of the current mainstream text classification methods are applicable to large-scale data samples. For the classification model in the low resource scenario, there will be overfitting problems on small samples, while in the e-commerce platform transaction process, there will be a situation where the amount of large-scale data cannot be reached. In order to fully extract the Semantic information of short text e-commerce data, a small sample e-commerce classification research based on prompt learning was proposed. Firstly, based on the pre trained language model BERT (Bidirectional Encoder Representations from Transformer) for feature vector representation, an e-commerce text prompt template is designed based on the idea of prompt learning. The text classification task is transformed into a cloze filling task, and the predicted classification label is obtained through model training. The editing distance is calculated to obtain the final label. Comparative experiments were conducted on different sample sizes, datasets, and models, and the experimental results showed that the proposed method had better results, verifying the effectiveness of the model and having certain practical value for commodity analysis on e-commerce platforms.