Wicaksono, Reo (2025) Exploration of hierarchical deep learning for automatic quality assessment of open-ended concept map propositions / Reo Wicaksono</p>. Masters thesis, Universitas Negeri Malang.
Full text not available from this repository.Abstract
Automatic assessment of open-ended concept maps is still inaccurate because each proposition contains hierarchical structures and highly varied language styles. This diversity makes it difficult for rule-based systems and conventional statistical models to understand the relationships between concepts especially with long propositions. This research utilizes hierarchical deep learning to address the complexity of both long and short propositions by comparing three architectures the Hierarchical Convolutional Neural Network and two recurrent models the Hierarchical Gated Recurrent Unit and the Hierarchical Long Short-Term Memory. The dataset includes 690 database concept propositions and 1 466 cybersecurity concept propositions. Class imbalance is addressed using SMOTE and class weighting to prevent model bias. Evaluation was conducted using accuracy precision recall F1 and Cohen s Kappa. Experiments showed that HCNN consistently outperformed the other models achieving 96% accuracy ( kappa 0.94) on relational database data and 95% accuracy ( kappa 0.92) on cybersecurity data. This high performance underscores the effectiveness of hierarchical convolutions in recognizing key n-grams alongside global context. HGRU ranked second while HLSTM was slightly below it. Both struggled to distinguish classes with similar semantic features especially in the more heterogeneous cyber security domain. These findings indicate that for automatic evaluation of open-ended concept maps the hierarchical convolutional approach is more adaptive than recurrent memory mechanisms. Hierarchical CNN is more appropriate because its multi-level convolutional filters can extract key n-grams while simultaneously summarizing local-global patterns in parallel thereby avoiding the long-term context loss constraints that often limit recurrent memory mechanisms in RNNs. This research not only provides a benchmark for the model but also opens up opportunities for integrating HCNN into e-learning systems to provide fast and consistent feedback for educators and learners.
| Item Type: | Thesis (Masters) |
|---|---|
| Divisions: | Fakultas Teknik (FT) > Departemen Teknik Elektro (TE) > S2 Teknik Elektro |
| Depositing User: | library UM |
| Date Deposited: | 19 Aug 2025 04:29 |
| Last Modified: | 09 Sep 2025 03:00 |
| URI: | http://repository.um.ac.id/id/eprint/390733 |
Actions (login required)
 |
View Item |