Answers are grounded in this paper and cite their sources.
Portuguese–Chinese legal translation requires stable terminology rendering, accurate clause-level interpretation, and reliable handling of legal concepts across different legal systems. Conventional neural machine translation models often show inconsistent term selection and weakened logical relations when processing long legal sentences, embedded conditions, and cross-clause references. This study proposes a terminology-aware Portuguese–Chinese legal NMT framework based on hybrid attention and dynamic legal terminology embedding. The framework introduces term type-aware positional encoding into the Transformer encoder to distinguish specialized legal terms from common lexical units. During decoding, a dynamic term-gating mechanism modulates a position-dependent attention bias using the current decoder context and retrieved legal-term vectors, while position-sensitive attention supports clause numbers, article references, and long-distance legal dependencies. Improved BM25 and cross-lingual BERT were used for candidate extraction and semantic-equivalence filtering, producing an experimental terminology database of approximately 800,000 aligned entries after conflict resolution and quality filtering. Locality-sensitive hashing and mixed-precision quantization were used to improve retrieval efficiency and reduce memory cost. Under the available 8:1:1 sentence-level partition, five random-seed runs yielded 31.4 ± 0.8 BLEU, 49.2 ± 1.2 TER, and 89.5 ± 0.5 Term F1. Relative to mBART-large, BLEU increased by 1.9 points (95% CI: 1.3–2.5; exact p < 0.001; Bonferroni-adjusted p = 0.002) and Term F1 increased by 11.3 points (95% CI: 10.5–12.1; exact and adjusted p < 0.001). The proposed system achieved 8200 words/s throughput and 218 ± 12 ms average full-translation latency in the standardized model-comparison setting. Because the retained provenance records do not permit retrospective verification of document-level splitting or complete test-set isolation from all terminology resources, the reported gains should be interpreted as conditional on the available evaluation partition rather than as a leakage-free document-level benchmark. The results nevertheless provide evidence that terminology-aware modelling and scalable terminology retrieval can improve automatic Portuguese–Chinese legal translation metrics, while broader document-level and expert legal validation remain necessary.
No takes yet. Share an insight, caveat, or question.
Song et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: