Terminology Alignment based on Multi-level Feature Fusion for Japanese Scientific and Technological Literature Terminology Translation

• Published in: Journal of Applied Science and Engineering Vol. 29; no. 2; pp. 465 - 473
• Main Author: Qi Fang
• Format: Journal Article
• Language: English
• Published: Tamkang University Press 2026
• Subjects: multi-level feature fusion; target pseudo terminology generation; terminology alignment
• DOI: 10.6180/jase.202602_29(2).0021
• ISSN: 2708-9967

Bilingual terminology alignment databases represent crucial resources in the field of natural language processing, holding significant value for multilingual applications such as cross-lingual information retrieval and machine translation. Bilingual terminology pairs are typically obtained through manual translation or automatic extraction from bilingual parallel corpora. However, manual translation requires domain-specific expertise and proves time-consuming and labor-intensive, while large-scale bilingual parallel corpora in specific domains remain scarce. To this end, a terminology alignment method based on multi-level feature fusion (TA-MFF) is proposed for Japanese scientific and technological literature terminology translation. First, a multi-engine collaborative generation mechanism is designed to produce target pseudo terminology candidates through parallel translations from heterogeneous machine translation systems, effectively expanding the coverage of potential translations while mitigating single-engine bias. Second, a hybrid feature extraction architecture is constructed by integrating Transformer’s multi-head attention with BiLSTM’s sequential modeling capabilities, where positional encoding is deliberately omitted to leverage BiLSTM’s inherent strength in capturing positional relationships, thereby enhancing context-aware feature representation. Third, an adaptive multi-level fusion strategy is developed through the synergistic combination of soft attention-based global interaction features and cosine similarity-based local interaction features, with trainable weights automatically balancing their respective contributions to achieve comprehensive semantic modeling. These innovations collectively address the critical challenges of translation ambiguity reduction, cross-lingual feature alignment, and multi-perspective similarity evaluation in Japanese scientific terminology alignment, ultimately improving both precision and robustness compared to conventional approaches.