Sistem Deep Question Generation Berbasis EBloom Menggunakan Arsitektur Encoder-Decoder

Aisyah, Nur (2026) Sistem Deep Question Generation Berbasis EBloom Menggunakan Arsitektur Encoder-Decoder. Masters thesis, Institut Teknologi Sepuluh Nopember.

Text 6025241004-Master_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (3MB) | Request a copy

Abstract

Evaluasi pembelajaran sangat penting untuk menilai pemahaman siswa, namun penyusunan instrumen pertanyaan secara manual dalam jumlah besar sering kali membebani guru. Meskipun teknologi Automatic Question Generation (AQG) berkembang pesat, mayoritas sistem masih menghasilkan shallow questions (73%), sedangkan deep questions untuk menstimulasi proses berpikir tingkat tinggi hanya dibahas sebesar 9%. Keterbatasan ini dipicu oleh ketergantungan pada taksonomi tunggal konvensional seperti Taksonomi Bloom yang memiliki kelemahan tumpang tindih interpretasi kata kunci dengan indeks Shannon Evenness yang tinggi (J' ≈ 0,66). Untuk mengatasi celah ini, penelitian ini mengusulkan sistem Deep Question Generation (DQG) berbasis kerangka kognitif eBloom (enriched Bloom) yang mengintegrasikan Taksonomi Bloom, Taksonomi Graesser dan Pearson, Teori Halliday, Teori John Dewey, serta Teori Scott. Kerangka eBloom digunakan sebagai panduan eksplisit untuk membangun knowledge base melalui model Large Language Model (LLM) yaitu Gemini-2.5-Pro. Penelitian ini memanfaatkan dataset mandiri dari buku teks IPS jenjang SMP dan SMA kelas VII-XII yang mencakup 46 bab dan 177 subbab (5.310 pertanyaan), serta dataset publik lintas domain EduProbe (1.000 pertanyaan). Validasi oleh tiga guru ahli menghasilkan nilai reliabilitas Fleiss' Kappa sebesar κ = 0,82 (almost perfect agreement), dengan penyesuaian label sebanyak 8% pada dataset IPS dan 20,2% pada EduProbe. Sistem DQG dirancang menggunakan arsitektur dual-channel encoder-decoder. Pada sisi encoder, fitur sekuensial diekstrak menggunakan Bi-GRU berbasis IndoBERT, sementara fitur relasional diproses paralel menggunakan Semantic Graph Encoder berbasis Graph Convolutional Network (GCN) melalui ekstraksi struktur kalimat SPOK. Kedua representasi digabungkan secara dinamis melalui lapisan gated attention fusion. Pada sisi decoder, model GPT-2 dioptimasi dengan menyuntikkan cognitive level embedding sebagai prefix untuk mengarahkan kedalaman berpikir token pertanyaan melalui mekanisme cross-attention. Hasil eksperimen menunjukkan model usulan (M10) secara konsisten mencapai performa terbaik. Secara kuantitatif pada dataset IPS, M10 meraih skor tertinggi dengan nilai BLEU-4 (0,0414), ROUGE-L (0,1585), METEOR (0,1572), dan BERTScore (0,7259). Secara kualitatif melalui penilaian skala Likert oleh lima guru ahli, model ini unggul signifikan pada aspek relevansi (90,0%), kesesuaian level kognitif (90,6%), dan bahasa (90,0%). Penelitian ini membuktikan kombinasi konteks sekuensial, graf semantik, dan panduan kognitif eksplisit mampu menghasilkan pertanyaan mendalam yang akurat secara kontekstual dan tepat secara pedagogis.
======================================================================================================================================
Learning evaluation is vital to assess student comprehension, but manually constructing large numbers of high-quality assessment instruments often imposes a heavy administrative burden on teachers. Despite the rapid growth of Automatic Question Generation (AQG) technology, existing systems are heavily biased toward shallow questions (73%), while deep questions designed to stimulate higher-order thinking skills comprise only 9%. This limitation is driven by the reliance on single conventional frameworks like Bloom's Taxonomy, which suffers from keyword interpretation overlap with a high Shannon Evenness index (J' ≈ 0,66). To bridge this gap, this study proposes a Deep Question Generation (DQG) system based on the eBloom (enriched Bloom) cognitive framework, which integrates Bloom's Taxonomy with the Graesser and Pearson Taxonomy, Halliday's Theory, John Dewey's Theory, and Scott's Theory. The eBloom framework provides explicit guidance to construct a structured knowledge base using the Gemini-2.5-Pro Large Language Model (LLM). This study utilizes a self-developed dataset from Social Studies (IPS) textbooks for junior and senior high schools spanning grades VII to XII across 46 chapters and 177 sub-chapters (5,310 questions) alongside the cross-domain public dataset EduProbe (1,000 questions). Validation by three expert teachers yielded a Fleiss' Kappa reliability value of κ = 0.82 (almost perfect agreement), adjusting labels for 8% of IPS data and 20.2% of EduProbe data. The proposed DQG system is implemented using a dual-channel encoder-decoder architecture. Within the encoder component, sequential features are extracted using an IndoBERT-based Bi-GRU, while relational features are processed in parallel using a Graph Convolutional Network (GCN) semantic graph encoder via subject-predicate-object-adverb (SPOK) extraction. These dual representations are dynamically integrated via a gated attention fusion layer. On the decoder side, a GPT-2 model is optimized by injecting cognitive level embeddings as a prefix to implicitly direct the cognitive depth of generated tokens via a cross-attention mechanism. Experimental results demonstrate that the proposed model (M10) consistently achieves superior performance over baseline and state-of-the-art models. Quantitatively on the IPS dataset, M10 achieved the highest scores with BLEU-4 (0.0414), ROUGE-L (0.1585), METEOR (0.1572), and BERTScore (0.7259). Qualitatively via Likert-scale evaluations by five expert teachers, the model significantly excelled in relevance (90.0%), cognitive level accuracy (90.6%), and language quality (90.0%). This research successfully validates that synthesizing sequential language context, graph semantic mapping, and explicit cognitive signals generates deep questions that are contextually accurate and pedagogically sound.

Item Type:	Thesis (Masters)
Uncontrolled Keywords:	Arsitektur Encoder-Decoder, Deep Question Generation, eBloom, Gated Attention Fusion, Knowledge Base, Deep Question Generation, eBloom, Encoder-Decoder Architecture, Gated Attention Fusion, Knowledge Base
Subjects:	Q Science > QA Mathematics > QA336 Artificial Intelligence Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science)
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55101-(S2) Master Thesis
Depositing User:	Nur Aisyah
Date Deposited:	29 Jul 2026 08:29
Last Modified:	29 Jul 2026 08:29
URI:	http://repository.its.ac.id/id/eprint/139724

Actions (login required)

View Item