Fusion-Align：一种自动字幕翻译中的融合对齐同步方法
Fusion-Align: A Fusion Alignment and Synchronization Method for Automatic Subtitle Translation
华松 1罗自强2
作者信息
- 1. 海南师范大学 信息科学技术学院，海口 571127
- 2. 海南师范大学 数据科学与智慧教育教育部重点实验室，海口 571127
- 折叠
摘要
[目的] 面向自动字幕和语音翻译及其语料库构建任务中原文字幕与译文字幕对齐及时间同步的问题，提出 Fusion-Align 融合对齐同步方法。[方法] Fusion-Align 以多特征融合与温度标定的对数似然比（LLR）评分器为核心，配合 DTW 对齐算法进行全局最优路径搜索，并通过子句级对齐再优化提升复杂长句的对齐质量。方法从句级与词级嵌入向量中提取 25 维特征（9 维句级语义特征、10 维词无序特征与 6 维词有序特征），经轻量 MLP 输出对数几率并通过温度标定得到可加 LLR 作为统一评分刻度。[结果] 在真实 TED 数据的所评样本上，融合对齐方式（B）的系统均分明显高于直译对齐（A）；配对自助法的置信区间完全为正，95\% 置信水平下支持“B 优于 A”的结论。[局限] 当 ASR 误识/断句漂移显著或跨域分布偏移较大时，刻度温度与阈值的自适应校准仍有空间；单调路径对大幅度重排仍有约束；跨域/跨语种的温标稳健性与特征转移尚待加强。[结论] Fusion-Align 把多源证据凝练为统一可加的 LLR 标尺，将“多特征融合→统一刻度（LLR）→单调全局搜索（DTW）→子句级对齐再优化”的流程闭环，为自动字幕翻译在真实 ASR 场景中的对齐问题提供了新的方法与技术路线。
Abstract
[Objective] To address the problem of aligning source subtitles and their translations and synchronizing timestamps in automatic subtitling/speech translation and corpus construction, we propose Fusion-Align.[Methods] Fusion-Align builds an additive log-likelihood-ratio (LLR) scoring model via multi-feature fusion with temperature calibration, and searches a global optimal monotonic alignment path with DTW; a clause-level re-optimization further refines alignments for long sentences. We extract 25 features from sentence- and word-level embeddings (9 sentence-semantic, 10 order-free lexical, and 6 order-aware lexical features), and feed them into a lightweight MLP to produce calibrated additive LLR scores.[Results] On real TED samples, the fused alignment setting (B) achieves a higher system score than the direct baseline (A); paired bootstrap confidence intervals are strictly positive, supporting B > A at the 95% confidence level.[Limitations] When ASR errors/segmentation drift are severe or domain shift is large, adaptive calibration of temperatures and thresholds remains to be improved; the monotonic path also constrains large-scale reordering; robustness and transfer across domains/languages need further study.[Conclusions] Fusion-Align closes the loop of feature fusion unified LLR scale monotonic global search (DTW) clause-level refinement, providing a robust and interpretable alignment route for realistic ASR-to-MT subtitle production.关键词
自动字幕翻译/特征融合/多层感知机/DTW对齐算法Key words
Automatic subtitle translation/Feature fusion/Multilayer perceptron/DTW alignment algorithm引用本文复制引用
华松,罗自强.Fusion-Align：一种自动字幕翻译中的融合对齐同步方法[EB/OL].(2025-12-26)[2025-12-27].https://chinaxiv.org/abs/202512.00239.学科分类
计算技术、计算机技术
评论