[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"detail-sidebar-cat-0-en-105":3,"doc-seo-273522-105":59,"doc-detail-273522-en":131},{"code":4,"msg":5,"data":6},0,"success",[7,13,18,23,28,33,38,43,48,51,55],{"id":8,"doc_module":4,"doc_module_name":9,"category_name":10,"show_sort_weight":11,"slug":12},1,"Document","Story & Novel",90,"story-novel",{"id":14,"doc_module":4,"doc_module_name":9,"category_name":15,"show_sort_weight":16,"slug":17},2,"Literature",80,"literature",{"id":19,"doc_module":4,"doc_module_name":9,"category_name":20,"show_sort_weight":21,"slug":22},4,"Exam",70,"exam",{"id":24,"doc_module":4,"doc_module_name":9,"category_name":25,"show_sort_weight":26,"slug":27},5,"Comic",60,"comic",{"id":29,"doc_module":4,"doc_module_name":9,"category_name":30,"show_sort_weight":31,"slug":32},6,"Technology",50,"technology",{"id":34,"doc_module":4,"doc_module_name":9,"category_name":35,"show_sort_weight":36,"slug":37},7,"Healthcare",40,"healthcare",{"id":39,"doc_module":4,"doc_module_name":9,"category_name":40,"show_sort_weight":41,"slug":42},8,"Research & Report",30,"research-report",{"id":44,"doc_module":4,"doc_module_name":9,"category_name":45,"show_sort_weight":46,"slug":47},9,"Religion & Spirituality",20,"religion-spirituality",{"id":46,"doc_module":4,"doc_module_name":9,"category_name":49,"show_sort_weight":46,"slug":50},"World Cup","world-cup",{"id":52,"doc_module":4,"doc_module_name":9,"category_name":53,"show_sort_weight":52,"slug":54},10,"Lifestyle","lifestyle",{"id":56,"doc_module":4,"doc_module_name":9,"category_name":57,"show_sort_weight":24,"slug":58},19,"General","general",{"code":4,"msg":60,"data":61},"ok",{"site_id":62,"language":63,"slug":64,"title":65,"keywords":66,"description":67,"schema_data":68,"social_meta":124,"head_meta":126,"extra_data":128,"updated_unix":130},105,"en","an-investigation-of-translationese-in-the-generations-of-multilingual-large-language-models","An Investigation of Translationese in the Generations of Multilingual Large Language Models","","Text translated from another language often retains linguistic cues known as translationese. Multilingual large language models (MLLMs) generate in many languages, but it remains unclear whether their outputs reflect internal translation processes and therefore exhibit translationese. This work tests two research questions on resemblance to translationese and differences from direct translation by using established translation indicators, classification models, feature-level variance analyses, and human annotations for German and Spanish.",{"@graph":69,"@context":123},[70,84,106],{"@type":71,"itemListElement":72},"BreadcrumbList",[73,77,79,82],{"item":74,"name":75,"@type":76,"position":8},"https://docshare.wps.com","Home","ListItem",{"item":78,"name":9,"@type":76,"position":14},"https://docshare.wps.com/document/",{"item":80,"name":40,"@type":76,"position":81},"https://docshare.wps.com/document/research-report/",3,{"item":83,"name":65,"@type":76,"position":19},"https://docshare.wps.com/document/an-investigation-of-translationese-in-the-generations-of-multilingual-large-language-models/273522/",{"url":83,"name":65,"@type":85,"image":86,"author":91,"headline":65,"publisher":94,"fileFormat":97,"inLanguage":63,"description":67,"dateModified":98,"datePublished":99,"encodingFormat":97,"isAccessibleForFree":100,"interactionStatistic":101},"DigitalDocument",{"url":87,"@type":88,"width":89,"height":90},"https://docshare.wps.com/thumbnails/an-investigation-of-translationese-in-the-generations-of-multilingual-large-language-models/273522.png","ImageObject",300,407,{"name":92,"@type":93},"Hazel","Person",{"url":74,"name":95,"@type":96},"DocShare","Organization","application/pdf","2026-10-30","2026-09-15",true,{"@type":102,"interactionType":103,"userInteractionCount":105},"InteractionCounter",{"@type":104},"ViewAction",15,{"@type":107,"mainEntity":108},"FAQPage",[109,115,119],{"name":110,"@type":111,"acceptedAnswer":112},"What is translationese and why is it important for evaluating multilingual LLM outputs?","Question",{"text":113,"@type":114},"Translationese refers to linguistic evidence that remains after text is translated. It matters because it can reveal whether model generations resemble translated text rather than truly native-language production.","Answer",{"name":116,"@type":111,"acceptedAnswer":117},"Which research questions does the investigation address?",{"text":118,"@type":114},"The study asks whether MLLM-generated text resembles translationese and how MLLM translationese differs from translationese produced via direct translation.",{"name":120,"@type":111,"acceptedAnswer":121},"How do the authors assess translationese in MLLM generations?",{"text":122,"@type":114},"They use established translated-text indicators, high-accuracy classification models, analyses of variance on linguistic features, and human annotations for selected languages (German and Spanish).","https://schema.org",{"og:url":83,"og:type":125,"og:title":65,"og:site_name":95,"og:description":67},"article",{"robots":127,"canonical":83},"index,follow",{"doc_id":129,"site_id":62},273522,1789471857,{"code":4,"msg":5,"data":132},{"doc_id":129,"user_id":133,"nickname":92,"user_avatar":134,"doc_module":4,"category_id":39,"category_name":40,"doc_title":65,"doc_description":67,"doc_content":135,"file_id":136,"file_url":137,"file_type":138,"file_size":139,"view_count":105,"is_deleted":4,"is_public":8,"is_downloadable":8,"audit_status":8,"page_count":140,"language":141,"language_code":63,"site_id":62,"html_lang":63,"table_of_contents":142,"faqs":143,"seo_title":144,"seo_description":67,"update_tm":130,"read_time":145,"preview_page_count":4},137441390410,"https://ap-avatar.wpscdn.com/avatar/2000252f4ab5702993?_k=1776741390130283984","arXiv : 2608 . 17399v 1 [cs .CL] 18 Aug 2026  \nAn Investigation of Translationese in the Generations of Multilingual Large Language Models  \nMaria Valentini1, Téa Wright2, Julisa Granados1, Eliana Colunga1 & Katharina von der Wense1,3  \n1University of Colorado Boulder  \n2University of California Berkeley  \n3Johannes Gutenberg University Mainz {[first.last}@colorado.edu](first.last}@colorado.edu)  \nAbstract  \nText which has been translated from another language tends to carry with it evidence of translation—hence, it is often referred to as translationese.  \nMultilingual large language models (MLLMs) generate text in a variety of languages. However, it is still unclear if MLLMs’ generations resemble internal translation (from English or, potentially, other languages) and, thus, result in translationese. Here, we ask the following research questions: (1) Does text generated by MLLMs resemble translationese? (2) How does translationese produced by MLLMs differ from translationese produced through direct translation? We leverage established indicators of translated text to evaluate text generated by state-of-the-art MLLMs in five languages, comparing to both non-translated and human-written baselines in order to isolate translationese from other kinds of interference. Through the use of high-accuracy classification models, analyses of variance on individual linguistic features, and the collection of human annotations for a subset of two languages (German and Spanish), we assess the translationese content of MLLM generations and examine the key features that distinguish MLLMgenerated text from typical translation-related interference.  \n1 Introduction  \nIt is well documented that large language models (LLMs) are becoming increasingly adept at generating text in non-English languages. Systems such as ChatGPT (OpenAI, 2022), while not explicitly designed for multilingual generation, have demonstrated some level of ability to generate text in non-English languages. Other models such as PaLM2 (Anil et al., 2023) are explicitly designed for multilingual use and claim improved fluency for many languages when compared to their counterparts which are primarily trained on English.  \nThe linguistic naturalness of LLM-generated text is an important component of its overall quality, highly relevant for human-facing applications such as designing language learning materials. However, it is often overshadowed by more traditionally important attributes, such as correctness, coherence, and relevance. While these are all critical for assessment, LLMs should ideally also mirror the naturalness of native speakers, regardless of language.  \nIn recent years, apparent multilingual proficiency in MLLMs has raised the question of whether they truly generate text in non-English languages in a linguistically native way or whether their outputs instead reflect translation from a dominant internal language (e.g., English) . Given the black-box nature of LLMs, it is difficult to know exactly what is going on behind the scenes when text is generated. However, recent work suggests that MLLMs may rely on English-centric internal representations, even when prompted in another language (Schut et al., 2025) .  \nHere, we take a different approach and focus on MLLM generations: we evaluate LLMs using externally observable phenomena typical for translated text, known as translationese.  \nThis has potential impact both in assessing LLM naturalness for human-facing applications and interpretability research. Fig. 1 shows a simple example of translationese.  \nIn this paper, we present several experiments, with the goal of answering the following research questions: (1) Does text generated by MLLMs resemble translationese? (2) How does translationese produced by MLLMs differ from translationese produced through direct translation? In our first experiment, we focus on the high-and medium-resource languages of English, German, Spanish, and Greek. We compare LLM-generated tex","cbCaiv6ozLYh1iPw","https://ap.wps.com/l/cbCaiv6ozLYh1iPw","pdf",398749,17,"English","# Abstract\n# Introduction\n# Related work\n## Translationese","[{\"question\":\"What is translationese and why is it important for evaluating multilingual LLM outputs?\",\"answer\":\"Translationese refers to linguistic evidence that remains after text is translated. It matters because it can reveal whether model generations resemble translated text rather than truly native-language production.\"},{\"question\":\"Which research questions does the investigation address?\",\"answer\":\"The study asks whether MLLM-generated text resembles translationese and how MLLM translationese differs from translationese produced via direct translation.\"},{\"question\":\"How do the authors assess translationese in MLLM generations?\",\"answer\":\"They use established translated-text indicators, high-accuracy classification models, analyses of variance on linguistic features, and human annotations for selected languages (German and Spanish).\"}]","An Investigation of Translationese in the Generations of Multilingual Large Language Models | PDF",43]