상세 보기
초록
Previous research on low-resource neural machine translation (NMT) has primarily focused on improving the translation performance, while the structural characteristics of the alignment distributions have received limited attention. This paper proposes a Complex-Attention Transformer (CAT) that extends self-attention in the encoder and decoder into the complex-valued domain, applies Hermitian operations to the decoder cross-attention mechanism, and uses CAT to analyze the structural characteristics of alignment distributions. Experiments on four low-resource language pairs showed that CAT increases the alignment entropy by 9.7%–35.6% compared with a parameter-matched real-valued model (RealMatch), while the translation quality exhibits no significant difference in most conditions. The Random Forest (RF) and SHapley Additive exPlanations (SHAP)-based structural analysis further indicates that whether the translation quality is preserved is more closely associated with the structural interactions among alignment-related features than with any single feature. Case analyses further showed that increased entropy reflects the structural expansion of the alignment search space rather than the degraded translation quality. These findings suggest that complex-valued attention can expand the representational scope of alignment distributions while preserving translation quality, and provide directions for the design of low-resource NMT models.
키워드
- 제목
- 저자원 신경기계번역에서 복소수 어텐션에 의한 정렬 확장과 품질 보존의 구조적 분석
- 제목 (타언어)
- Structural Analysis of Alignment Expansion and Quality Preservation via Complex-Valued Attention in Low-Resource NMT
- 저자
- 송선호; 김창근; 김봉기; 김현주
- 발행일
- 2026-10
- 유형
- Y
- 저널명
- 한국산학기술학회논문지
- 권
- 27
- 호
- 10
- 페이지
- 15 ~ 25
- 언어
- KOR
- 출판사
- 한국산학기술학회
- 발행국가
- 대한민국
- 분량
- 11 페이지
- ISSN
- E 2288-4688
P 1975-4701