Evaluation of Structural Hyperparameters for Text Classification with LSTM Networks

In natural language processing, most problems can be interpreted as text classification tasks, which makes this type of task a central one. A natural subdivision of more complex types of text classification can be made according to whether the classification is multilabel or multiclass, where both of these can be tackled with either a multiclass classifier or with a combination of several binary classifiers. An example of the problem which offers a natural way of comparing multiclass and binary classification on the same data, which makes the results comparable, is a classification of text author MBTI personality type. The dataset used is PersonalityCafe MBTI. We focus our comparison on structural hyperparameters, which are the hyperparameters pertaining to network structure. Structural hyperparameters are necessary for specifying the network itself, which makes them a primary concern in its construction. The hyperparameters investigated in this paper are the number of hidden layers and layer size. Through a number of experiments, we demonstrate the choice of hyperparameters and conclude with general hyperparameter selection recommendations based on our results.