An Industrial Short Text Classification Method Based on Large Language Model and Knowledge Base

Text classification is a common task in natural language processing tasks. However, in reality, such as industrial datasets, there are numerous short texts that are ignored by existing models, in which important information is inevitably missed due to the short word count. Recent studies tend to enrich the features in short text classification by introducing conceptual information. However, these simple conceptual information is also relatively sparse, with limited feature augment. In this paper, we propose Input Embedding, Knowledge Base Retrieval, Large Language Model Text Augmentation, and Text Encoding for the task of short text classification (IKLT). The four modules utilize the current well-performing forms of Large Language Model and Retrieval Augmentation to further enrich the feature information and have interpretability. Moreover, in addition to some public datasets we establish an industrial short text dataset for comparison experiments. The experimental results show that our proposed framework for short text classification achieves improved results on all five datasets.