Generative artificial intelligence (GenAI) has advanced machine translation (MT) significantly, enabling translation and revision through prompts in AI systems like ChatGPT. While neural machine translation (NMT) is known for accuracy and fluency, particularly in managing lexical inversion, improving readability, and enhancing textual coherence, its shortcomings remain. This study addresses the underexplored area of GenAIT's linguistic features and (dis)advantages compared to human translation (HT). Focusing on scientific texts due to their crucial role in knowledge dissemination, the study compares translations by ChatGPT 3.5 and 19 Master of Translation and Interpreting (MTI) students in China. It hypothesizes distinct linguistic features exist between GenAIT and HT, and their strengths can be integrated for superior translations.

Existing research comparing machine translation (MT) and human translation (HT) linguistic features falls into two categories: translation studies and linguistic analysis. Translation studies research focuses on translation universals, quality, and translationese. Studies have explored simplification, explicitation, and convergence in MT versus HT, alongside quality metrics like BLEU and TER. Linguistic analysis research examines lexical analysis, morphologically complex words, cohesion, syntactic features, and figurative language. Studies note differences in lexical tightness, diversity, consistency, use of cohesive devices, passive voice, tense, sentence structure, and handling of figurative language. However, gaps remain: limited research on GenAIT-HT linguistic comparison in scientific texts, especially those involving Chinese; a primary focus on Indo-European languages. This study aims to address these gaps by examining the linguistic characteristics of GenAIT and HT in Chinese-English scientific text translation.

This empirical study involved 19 MTI students translating two assignments (English-to-Chinese and Chinese-to-English). This study focused solely on the English-to-Chinese translations, leveraging the translators' strength in translating into their native language. Participants used only printed dictionaries, ensuring a standardized process. The English source texts (ST) were adapted from standard scientific and technological translation teaching materials to ensure appropriateness and difficulty. The length was chosen to align with the workload of the Level-2 China Accreditation Test for Translators and Interpreters (CATTI), targeting 243 words. ChatGPT 3.5 generated GenAIT using the prompt "please translate the following passage into Chinese." A Reference Translated Text (RTT) from the same source material was also included. Data analysis utilized Wordless 2.3.0, Corpus WordParser 3.0, CLAWS-5, and AntConc 4.1.2 for automatic data collection (tokens, types, STTR, CS, SLT, POS). Terminology, passive voice (PV), and subordinate clauses were analyzed manually. The analysis framework encompassed lexical features (tokens, types, STTR, terminology, POS – focusing on nouns, adjectives, numerals, and conjunctions) and syntactic features (sentence count (CS), average sentence length in tokens (SLT), passive voice structures, and subordinate clauses – focusing on object and attributive clauses).

At the lexical level, HT (mean) had more tokens and types but lower STTR (Standardized Type-Token Ratio) compared to GenAIT. GenAIT showed higher accuracy in terminology translation. HT (mean) had more nouns, adjectives, numerals, and conjunctions than GenAIT. However, some HT showed fewer tokens and types than the ST. At the syntactic level, HT (mean) exhibited more sentences but shorter average sentence length than GenAIT. Human translators frequently shifted passive voice to active voice, particularly in converting implicit passive structures, while GenAIT maintained explicit passive constructions but often missed implicit ones. HT demonstrated superior skills in breaking down long sentences, evident in their handling of object and attributive clauses, often segmenting complex sentences into simpler structures. GenAIT, although not adept at splitting long sentences, sometimes improved on the logical flow by adding connectors.

The observed differences between HT and GenAIT reflect both the strengths and weaknesses of each approach. HT’s longer texts and lower STTR might stem from simplification strategies, linguistic differences between languages, and individual translator familiarity with scientific texts. GenAIT’s higher STTR suggests richer lexical diversity and proficiency in handling scientific style. Differences in passive voice translation relate to translation techniques and language-specific structures. HT's sentence segmentation skills highlight human cognitive abilities in managing complexity. GenAIT's occasional improvements in logical flow reflect its potential in handling specific linguistic aspects. The findings highlight the complementarity of GenAIT and HT. GenAIT offers a wider vocabulary and background knowledge, while HT excels in handling complex sentence structures and subtle nuances.

This study reveals distinct lexical and syntactic characteristics of GenAIT and HT in scientific texts. HT produces longer texts with lower lexical diversity but excels in sentence segmentation and passive voice transformation. GenAIT exhibits higher accuracy in terminology and sometimes enhances logical connections. The complementary nature of both suggests a collaborative approach, leveraging GenAIT’s vocabulary and HT’s nuanced understanding of sentence structure. Future research could explore other language pairs and text genres, and investigate the impact of prompt engineering on GenAIT performance.

While the study provides valuable insights, limitations include the relatively small sample size and focus on a specific text type and language pair. The analytical framework, while comprehensive, might not capture all relevant linguistic features. Future research should address these limitations by expanding the sample size, exploring diverse text genres and language pairs, and incorporating additional parameters into the analysis framework.