To apply machine-learning technology in facilitating the discussion, helping people reach a consensus and other fields, classifying discussion contents according to the labels of the issue-based information system is critical and inevitable. However, the accuracy of classifying and predicting the discussion contents written in English is not satisfying right now. In this paper, two different pre-trained language models are fine-tuned to classify and predict IBIS labels of English sentences that were picked from two real discussion records. Based on the baseline of accuracy, several improvement methods are raised and taken. By comparing those results, the significant factors affecting the accuracy will be found to raise more detailed methods to improve the results in further research.