This dissertation examines consecutive interpreting in single-speaker public discourse within the political and diplomatic domain, with particular attention to how human interpreting actions emerge through variations in contextual mechanisms that shap...
This dissertation examines consecutive interpreting in single-speaker public discourse within the political and diplomatic domain, with particular attention to how human interpreting actions emerge through variations in contextual mechanisms that shape textual production. Rather than treating interpreting solely as a process of linguistic transfer, this study approaches interpreting as an activity in which interpreters actively reorganize discourse under situational, temporal, and audience-related constraints. These patterns are examined in comparison with outputs produced by neural machine translation (NMT) systems and large language model (LLM)–based translation.
Empirical data were collected through a series of controlled experiments involving master’s students in their final semester, who performed consecutive interpretations of presidential speeches and international forum presentations. For the same source texts, NMT and LLM translation outputs were generated and analyzed. In order to further distinguish interpreting from post-editing behavior, the same human participants were also asked to post-edit NMT outputs, allowing for a direct comparison between interpreting agency and post-editing intervention.
The analysis indicates that human interpreters do not simply reproduce source-text structures. Instead, they frequently adjust sentence boundaries, reorganize information, and reinforce levels of formality in response to the communicative situation and the intended audience. At the same time, certain elements of the source text are occasionally left unrealized. These omissions, however, are better understood as the result of selective restructuring under time pressure rather than as straightforward errors. NMT outputs, by contrast, show persistent features of translationese, including repetitive use of identical connectives such as y, which often reduces textual readability. Difficulties were also observed in cases requiring inferential recovery of ellipted elements, sometimes leading to distorted representations of the original information.
LLM-based translations demonstrate more flexibility than NMT in adjusting sentence boundaries and performing inference, although this tendency appears to be strongly influenced by prompt design. While recent advances in large language models have expanded the scope of machine translation, this study proceeds from the assumption that such systems remain fundamentally constrained in handling contextual risk in political discourse. What is particularly noteworthy is that LLM outputs occasionally introduce information absent from the source text or remove parts of the speaker’s utterance without explicit justification. From the perspective of political and diplomatic interpreting, such behavior raises serious concerns related to risk management.
By foregrounding the role of contextual mechanisms in consecutive interpreting, this study clarifies qualitative differences between human interpreting, machine translation, and post-editing practices. Rather than evaluating machine translation performance per se, this dissertation uses machine-generated outputs as a contrastive lens through which the nature of human interpreting agency becomes visible. It further argues that while human–machine collaboration in public discourse settings may become increasingly common, such collaboration requires a cautious framework supported by clearly defined and domain-sensitive guidelines.