Journal Paper Published

Abstract

The emergence of language technologies has positioned machine interpreting (MI) as a scalable solution to real-time multilingual communication, necessitating systematic examinations of its linguistic characteristics compared to those of human interpreting (HI) in order to foster more human-like outputs. Whereas initial studies have investigated various linguistic dimensions, syntactic complexity – a key indicator of linguistic sophistication and processing – remains underexplored in MI–HI comparisons. This study sought to bridge this gap by leveraging machine-learning classifiers to differentiate MI and HI based on multidimensional syntactic complexity metrics. We compiled a comparable Chinese-to-English corpus from government press conferences featuring HI renditions by professional interpreters and MI outputs generated by the iFlytek platform from identical source speeches. Ten machine-learning algorithms were trained on selected metrics across five dimensions, with ensemble models developed from top-performing classifiers and SHapley Additive exPlanations (SHAP) applied to quantify the importance of features. The results show that the equal-weighted ensemble of SVM, GB, and MLP yielded optimal discriminative performance (AUC = 84.29%, Accuracy = 76.15%). SHAP analysis revealed a distinct hierarchy of feature salience. Specifically, MI outputs exhibited “additive complexity”, which reflects sequential, modular processing within cascading architectures, whereas HI demonstrated “integrative complexity” derived from conceptually mediated processing that balances cognitive constraints against communicative goals. These findings advance our theoretical understanding of computational versus human language-processing mechanisms and provide empirical foundations for developing more naturalistic MI systems, thereby contributing to the emerging field exploring the interplay between interpreting and technology.

Keywords

ensemble model, human interpreting, machine interpreting, machine-learning classification, SHapley Additive exPlanations, syntactic complexity