Quad-Faceted Feature-Based Graph Network for Domain-Agnostic Text Classification to Enhance Learning Effectiveness

Enhancing learning effectiveness requires one to define suitable learning outcomes and align assessment constructs with these outcomes. We present a quad-faceted feature-based graph network to classify assessment texts into domain-agnostic class labels more accurately. The proposed model incorporates four complementary graphs (syntactic, semantic, sequential, and topical) with observable and latent node types and unique edge weight computations that are dependent on node properties to extract unique features from a given text. The purpose of incorporating syntactic information is to consider the dependency parsing between word nodes, while the semantic information is to provide the algorithm with contextual similarity between phrase nodes that are more effective than words in encapsulating the meaning of a text. The sequential graph is applied to regular expression nodes that contribute to a domain-agnostic class label, while the topical graph identifies topics that are convergent to each other based on their distributions. As opposed to existing techniques that construct graphs solely based on word nodes, the proposed model exploits the benefits of term weighting, nested phrases, regular expressions, and topic modeling to develop a diverse heterogeneous architecture for text classification. We evaluate the classification performance on questions with different class labels such as cognitive complexities, reasoning capabilities, and question types, as well as longer documents. Experiment results show that the proposed model outperforms in terms of macroaverage F1 score when compared with existing deep learning techniques. We also demonstrate the application of the classification model to understand learners℉ attitudes via an empirical study in a workplace-learning environment.