A Method Comparison on Multi-Label Questions Classification for Assessment-Based Personalised Scaffolding Adaptive Learning Path

Abstract

Classification of the topic of a question item is one of the fundamental problems in e-learning systems. Unlike single-label classification, the multi-label classification method simultaneously predicts more than one-class label. This research is a series of process development for a Personal Diagnostic system based on assessment. This system needs annotated question bank because multi-label question items can be used to build a Concept Effect Relationship (CER). The purpose of building CER is to track the failed concept of students who fail the formative tests. Hence, there is necessary in looking for a multi-label question classification method. Therefore, this paper compares several multi-label classification methods in determining subject topics associated with questions in a formative test question bank. This study investigates the non-neural-based and neural-based multi-label classification. The test results for the non-neural show that Term Frequency– Inverse Document Frequency (TF-IDF) with Random Forest classifier produces the best hamming loss value (16,3%) while on neural, TF-IDF with convolutional neural network (CNN) produces a hamming loss value (21,2%) that is better than Long Short Term Memory (LSTM).