Domain-specific Topic Modeling Configurations for Explainable Document-level Translation Evaluation

Published: 5 January 2026| Version 3 | DOI: 10.17632/k65v3vxwhb.3
Contributors:
Hyeokmin Lee, Youngkyu Kim, Byounghyun Yoo

Description

1. LSA_elbow_points.pdf: Contains per-domain elbow points derived from singular value decay curves (up to 30 components). Used to determine the optimal number of LSA topics for each domain. 2. LSA_full_singular_values.pdf: Provides complete singular value distributions for all domains and both languages (Korean/English). Shows variance contribution across components and supports dimensionality analysis. 3. LSA_max_features.pdf: Shows max-feature sensitivity tests across domains to determine the optimal TF-IDF vocabulary size. Includes comparisons of 500/1000/10000 features and their impact on stability. 4. LDA_coherence.pdf: Shows per-domain LDA coherence (c_v) across 1–30 topics and marks the selected optimal topic count k with a red dashed line for choosing the number of LDA topics.

Files

Institutions

  • Korea Institute of Science and Technology

Categories

English, Natural Language Processing, Machine Translation, Korean Language

Licence