A Dual-Channel Transformer Model for Translationese Detection Integrating Syntactic and Lexical Features

With the growing use of neural machine translation and LLM-based translation, translationese detection has become important for assessing translation naturalness and source-language interference. Existing methods often rely on explicit features or unified pretrained representations, giving limited attention to the joint modeling of lexical deviation, syntactic transfer, and local evidence. This paper proposes a Transformer-based dual-channel framework that separately encodes lexical and syntactic information, models their interaction through cross-channel fusion, and introduces local auxiliary discrimination to support sentence-level prediction. Experiments on the WMT Chinese news test set and the UNPC Chinese formal corpus show that the proposed method improves performance over the strongest syntax-enhanced baseline on WMT and achieves comparable results on UNPC. The framework demonstrates stable cross-domain performance and can support machine translation filtering, post-editing assistance, and generated-text style identification.