Pembobotan adaptif fungsi objektif W2VPred berbasis Joint Maximum a Posteriori pada Word Embedding multi-domain

Maulana, Dika Putra (2026) Pembobotan adaptif fungsi objektif W2VPred berbasis Joint Maximum a Posteriori pada Word Embedding multi-domain. Sarjana thesis, UIN Sunan Gunung Djati Bandung.

Preview	Text 1_cover.pdf Download (97kB) | Preview
Preview	Text 2_abstrak.pdf Download (159kB) | Preview
Preview	Text 3_skbebasplagiarism.pdf Download (192kB) | Preview
Preview	Text 4_daftarisi.pdf Download (116kB) | Preview
Preview	Text 5_bab1.pdf Download (188kB) | Preview
	Text 6_bab2.pdf Restricted to Registered users only Download (400kB) | Request a copy
	Text 7_bab3.pdf Restricted to Registered users only Download (751kB) | Request a copy
	Text 8_bab4.pdf Restricted to Registered users only Download (4MB) | Request a copy
	Text 9_bab5.pdf Restricted to Registered users only Download (255kB) | Request a copy
	Text 10_daftarpustaka.pdf Restricted to Registered users only Download (116kB) | Request a copy
	Text 11_lampiran.pdf Restricted to Repository staff only Download (422kB) | Request a copy

Abstract

INDONESIA : Model W2VPred mempelajari representasi kata multi-domain dan struktur kedekatan antar domain secara simultan melalui tiga komponen fungsi objektif: Fidelity, Structure, dan Regularization. Namun, penyeimbangan ketiganya bergantung pada koefisien statis yang dicari melalui grid search yang mahal secara komputasi. Penelitian ini mengusulkan W2VPred-JMAP, sebuah reformulasi probabilistik berbasis Joint Maximum A Posteriori (JMAP) yang memanfaatkan estimasi ketidakpastian tugas (task uncertainty) untuk menggantikan bobot statis. Fidelity dimodelkan sebagai likelihood Gaussian, sementara Structure dan Regularization diformulasikan sebagai prior Gibbs, di mana parameter ketidakpastian setiap distribusi dipelajari secara dinamis sebagai bobot adaptif terhadap besaran galat masing-masing komponen. Eksperimen pada korpus WikiFoS (16 disiplin ilmu) dan New York Times (20 tahun) menunjukkan bahwa W2VPred-JMAP menyamai akurasi analogi baseline terbaik (69,05% pada Top-10) dan mereduksi kebutuhan grid search sebanyak 25 kombinasi (λ, τ) menjadi satu parameter skalar (ηs). Model ini terbukti sama efektifnya dengan baseline dalam membedakan makna kata polisemik lintas domain tanpa mengorbankan representasi kata, serta meningkatkan prediksi struktur (Recall@3) dari 47,92% menjadi 50,00%. ENGLISH : The W2VPred model simultaneously learns multi-domain word representations and inter-domain affinity structures through three objective function components: Fidelity, Structure, and Regularization. However, balancing them relies on static coefficients searched through computationally expensive grid search. This study proposes W2VPred-JMAP, a probabilistic reformulation based on Joint Maximum A Posteriori (JMAP) that leverages task uncertainty estimation to replace static weights. Fidelity is modeled as a Gaussian likelihood, while Structure and Regularization are formulated as Gibbs priors, where the uncertainty parameter of each distribution is dynamically learned as an adaptive weight relative to the error magnitude of each component. Experiments on the WikiFoS (16 disciplines) and New York Times (20 years) corpora show that W2VPred-JMAP matches the best baseline analogy accuracy (69.05% at Top-10) and reduces the grid search requirement of 25 combinations (λ, τ ) to a single scalar parameter (ηs). This model is proven to be as effective as the baseline in distinguishing the meaning of polysemic words across domains without sacrificing word representation, and improves structure prediction (Recall@3) from 47.92% to 50.00%.

Actions (login required)

View Item