I think we would need more information about the model architecture you are using and the features you engineered for training.
Something I like to do when starting to work on text classification problems is to gradually increase the complexity of the architecture and start from simple approaches.
So if you are using
tf-idf features and some sort of probabilistic model, such as
logistic regression you may have a look at libraries such as Lime or eli5 that can help you make sense and explain the prediction of your classification model.
In your case you may then look at the top-features of the misclassified text and increase your understanding of why the model derived at its decision.