Classification Method for Ultra-Short Texts via Large language Model-Based Semantic Expansion

With advancements in natural language processing, deep learning models have been widely applied to diverse classification tasks. Although pre-trained models like BERT demonstrate exceptional performance in handling standard NLP tasks, they underperform in classifying ultra-short texts due to their reliance on complete syntactic structures and rich contextual information. This paper proposes a semantic expansion method for ultra-short texts, which transforms sparse inputs into context-enriched sentences by leveraging large language models and rule-based sentence construction. Specifically, we integrate ingredient lists, cooking methods, and flavor profiles to generate both standard syntactic sentences and semi-structured representations, thereby converting ultra-short text classification into a sentence classification task with augmented contextual signals. Experiments on a dataset of 4,000 recipe texts using the fine-tuned BERT model demonstrate significant improvements across all evaluation metrics.