# Multiclass text classification with insufficient dataset

I am working on a multi class text classification dataset which has 2719 data (only training set is available test set is not available) and 256 classes. Sentence length of each data is very short, average sentence length is 5 and many data has only single word. Issue I am facing is out of 256 classes 81 classes have only single data. I have tried Naive bayes with TF-IDF but the accuracy is very poor. Could some one suggest good approach to proceed ?

Example:

• It seems that the classes are hierarchical. For example, make me a sandwich and are you male or female both fall under TM_PQ. (Maybe TM is a broader class). Let's follow with an example - if I ask are you male or what sex are you should it fall under TM_PQ_Are_You_Male_Or_Female? Or should it fall under a new class? For now it seems that there is a basic class, and you attach the sentence itself to make it a specific class. – tmrlvi Jul 9 '18 at 8:57