We propose an inference method for complex professional texts with computational expressions. We use the expert rules to locate and rewrite the expressions. We adopt the pre-trained language model as the initial model and select the high quality samples from historical dataset for pre-training the model. The positive samples in the prompt are retrieved from the high-quality data that are similar to the inferred instance and the negative ones are generated by expert rules. For the target task, we fine-tune the model with the given few samples. Experimental results show that the performance of this method outperforms baselines and is applicable to low-resource scenarios.