AI翻译中的信息结构处理：问题与展望

Information Structure Processing in AI Translation: Problems and Prospects

摘要: 信息结构是语言组织信息、突出交际重点并引导读者理解的重要机制，主要涉及话题–述题、已知信息–新信息、焦点–背景等关系。在人工智能翻译中，译文质量不仅取决于词汇和句法层面的对应，还取决于原文信息结构能否得到适当重构。已有篇章级神经机器翻译通过分层注意力和上下文建模改善语篇处理，大语言模型翻译研究也进一步探索了长上下文条件下的篇章翻译。因此，当前问题不宜概括为模型普遍缺少跨句处理能力，而应聚焦于上下文内容能否稳定转化为话题延续、信息状态识别和焦点功能保持。本文在梳理相关理论与技术研究的基础上，分析语序调整、焦点识别和跨语言标记转换中的潜在偏差，提出面向信息结构的改进框架：将信息状态、焦点类型和共指关系形式化为分层标注与可计算特征，通过上下文编码、多任务学习和对比损失增强功能保持，并结合语篇专项测试与人工后编辑进行评价。上述方案旨在提供可检验的研究路径，其有效性仍需通过跨语言、跨体裁实验验证。

Abstract: Information structure organizes discourse, highlights communicative emphasis, and guides interpretation through relations such as topic-comment, given-new, and focus-background. In AI translation, translation quality depends not only on lexical and syntactic correspondence but also on the appropriate reconstruction of these relations. Existing document-level neural machine translation research has improved discourse processing through hierarchical attention and contextual modeling, while research on large language models has further explored document translation with extended context. The central issue is therefore not a general absence of cross-sentence processing, but whether contextual information can be reliably used to preserve topic continuity, information status, and focus function. Drawing on linguistic theories and relevant technical studies, this paper examines potential shifts arising from word-order adjustment, focus identification, and cross-linguistic differences in information-structural marking. It proposes an improved framework oriented towards information structure that formalizes information status, focus types, and coreference relations as layered annotations and computational features, integrates them through contextual encoding and multi-task learning, and introduces a contrastive objective for functional preservation. Discourse-specific testing and human post-editing are also incorporated into the evaluation framework. These proposals constitute testable research directions rather than established empirical findings, and their effectiveness requires validation across languages and genres.