With the rise of large language models, their remarkable in-context learning capacity has attracted increasing attention. When selecting examples for in-context learning, if it is possible to choose examples from the candidate set that are similar to the new query, the model’s reasoning performance may be better. Existing methods typically retrieve relevant context examples based on the semantic similarity between sentences. At this point, if two sentences share a significant proportion of words, their sentence representations will exhibit a strong correlation, potentially leading to inappropriate retrieval. In the table question answering task, to alleviate the aforementioned issue, we introduce a more detailed classification criterion that categorizes questions based on their inquiry style. Our approach was tested on the WikiTableQuestions dataset, achieving an accuracy of 65.6%. Compared to methods that utilize conventional sentence representations for classification and state-of-the-art pretraining with fine-tuning, our approach achieves improvements of 5.8% and 2.8% in accuracy, respectively. We will publish all source codes of this work on GitHub for further research explorations.