Bridging Language Gaps in Open-Source Documentation with Large-Language-Model Translation 1 st Elijah Kayode Adejumo, 2 nd Brittany Johnson, 3 rd Mariam Guizani Abstract While open source communities attract diverse contributors across the globe, only a few open-source software repositories provide essential documentation, such as ReadMe or CONTRIBUTING files, in languages other than English. Recently, large language models (LLMs) have demonstrated remarkable capabilities in a variety of software engineering tasks, even though they are not explicitly designed for such purposes. We have also seen advances in the use of LLMs for translations in other domains and contexts. Despite this progress, little is known regarding the capabilities of LLMs in translating open-source technical documentation, which is often a mixture of natural language, code, URLs, and markdown formatting. To better understand the need and potential for LLMs to support translation of technical documentation in open source, we evaluated community activity regarding translations along with English-to-German translations of 50 ReadMe files from OpenAI’s ChatGPT 4 and Anthropic’s Claude. We found scarce translation activity, most of which was found in larger repositories, community-driven and opportunistic in nature. A comparison of LLM performance as translators and evaluators of technical documentation suggests LLMs can provide accurate translations. However, an analysis of resulting translations found fidelity challenges: both models struggled to preserve structural components (e.g., hyperlinks) and exhibited formatting inconsistencies. These findings highlight both the promise and the challenges of LLM-assisted documentation internationalization. As a first step towards translation-aware continuous integration pipelines, we introduce TRIFID , an early-stage translation fidelity scoring framework that automatically checks how well translations preserve code, links, and formatting. Our efforts provide a foundation towards automated LLM-driven support for creating and maintaining open source documentation. I Introduction & Background Users, contributors, and maintainers all play an integral part in the success and sustainability of open source innovation, which means providing valuable support for all roles is beneficial. Written documentation serves as a fundamental source of support in open source projects. One vital piece of documentation in open source repositories is the ReadMe file usually in markdown format [ 1 ] , which plays a vital role in introducing new users and potential contributors to the software project, making its content essential [ 2 ] . The importance of the ReadMe file has attracted attention from researchers who evaluate how these files are created [ 3 ] , how useful they are to users and contributors, and which techniques maintainers can adopt to improve them. Recent findings [ 4 , 5 , 2 ] have shown a positive correlation between a repository’s ReadMe file content and various repository meta-characteristics such as popularity, development progress, and community size, emphasizing the importance of high-quality, accessible ReadMe files across repositories. Another study [ 6 ] proposed the use of transformers to summarize extremely long ReadMe files, thereby supporting better comprehension. While the ability to leverage the ReadMe , and other documentation, is important and valuable to all parties, studies suggest existing documentation may not be accessible to diverse audiences [ 7 , 8 ] ,also, studies [ 9 , 10 ] revealed that lack of documentation in preferred languages presents contributor challenges and that contributors who are not confident in English are more likely to report struggles in following discussions. This accessibility challenge could become even more problematic as a recent report [ 11 ] from GitHub reveals a growing contribution index from non-English speaking countries all over the globe, further emphasizing the need to understand and improve the current state of documentation accessibility. In recent years, large language models (LLMs) have demonstrated significant strengths in supporting various aspects of software engineering [ 12 , 13 , 14 ] . A recent mining study revealed the use of ChatGPT to improve ReadMe files and other relevant documentation [ 15 ] . Prior work has also indicated the potential for LLMs to improve, simplify, and summarize open source documentation to reduce information overload and increase the readability for contributors [ 8 , 16 , 15 ] . Moreover, LLM-driven translation and quality estimation have demonstrated state-of-art performance outside of computing [ 17 , 18 , 19 , 20 , 21 , 22 , 23 ] ; in fact, prior work suggests that with proper prompting strategies LLMs can achieve translation quality comparable to that of human experts [ 24 ] . Given LLMs’ joint training on natural-language and code corpora, we hypothesize that modern LLMs are better suited than conventional machine translation pipelines to handle the mixed Markdown + code + URL structure found in technical documentation. Although all of these insights are promising, there is little evidence to support the ability of LLMs to be used to translate technical documents such as ReadMe files. We investigated documentation translation contributions in open source ReadMe files, categorized repositories by scale and mined pull requests and issues with title ”Translate Readme”. We found more than 500 translation requests related to ReadMe files and they also reveal community-driven efforts during events such as Hacktoberfest [ 25 ] . However, many contributions were rejected and the ones merged could not keep pace with English updates. We also sampled 50 ReadMe files from large repositories ( >10000 stars/forks) to evaluate ChatGPT and Claude’s English-to-German translation. For this evaluation, we introduced TriFid , an early stage framework for assessing preservation of code, URLs, and markdown formatting. Our results utilizing TriFid interestingly revealed preservation challenges: both LLMs struggled with maintaining URLs and markdown formatting and code snippets. Moving towards this vision, we will: 1. Fine-tune LLMs with structure aware objectives that explicitly penalize corruption of code, links, and markdown; 2. Optimize TriFid into a feedback loop that scores the translation from the fine-tuned models. 3. Validate the results with native human experts and professional technical translators to close the gap between automated fidelity and real-world usability. II Research Questions RQ1: What patterns of translation contribution in open-source repositories reveal the global significance for multilingual technical documentation, and how effective and sustainable are these contributions? RQ2: How effective are LLMs technologies ChatGPT and Claude in translating technical documentation such as README files with respect to technical accuracy, preservation of code/markdown formatting? III Methodology Our initial focus was on identifying translation patterns across open-source software repositories and evaluating LLMs’ translation capabilities for technical documentation. To achieve this, we used the following methodology: III-A Data Set To reveal translation activities across repositories, we analyzed repositories at three scales based on stars and forks, as these metrics typically represent community engagement and project adoption [ 26 , 27 ] . They were classified into small-scale, medium-scale, and large-scale using the following criteria: i Small-Scale: 100 - 5,000 stars and forks ii Medium-Scale: 5,000 - 10,000 stars and forks iii Large-Scale: Stars and forks greater than 10,000 Given the importance of the ReadMe file [ 28 ] , we considered it a strong foundation for understanding translation issues and contribution patterns. For each repository category (small-scale, medium-scale, and large-scale), we searched for and extracted pull requests and issues whose titles contained “translate README” using a keyword-based approach [ 29 ] . This allowed us to assess the current state of translation contributions and issues raised across different repository categories. We then analyzed the timestamps of these pull requests to understand contribution trends and identify when translation activity peaks, providing insights into the periods or seasons during which translation demand is most likely to surge. To evaluate how effective and sustainable these translation contributions are, we examined whether the pull requests were actually adopted (merged) and, if merged, how well they kept up with their rapidly evolving English source. We compared the rate of change in the English and translated versions 180 days (six months) after each pull request was merged. III-B LLM technologies for documentation translation To explore the feasibility of LLM-powered technical documentation translation, we selected the top 50 ReadMe files from large-scale repositories, as these projects are widely adopted by developers worldwide [ 27 ] . III-B 1 Translation Process As a first step, we focused on translation from English to German based on a recent report from GitHub [ 30 ] that indicated Germany ranks among the top three contributors by region outside of the US. In addition, various LLM translation and evaluation studies have also evaluated this language translation, indicating that German is a high-resource language [ 31 , 32 , 33 ] . We provided the extracted markdown files to the LLMs (ChatGPT-4o and Claude Sonnet) using the following prompt: “You are an expert translator, please translate the following Markdown content from English to German. Preserve code blocks, links and markdown format exactly as they are.” In recent years, LLM prompting has been evolving, with studies [ 34 ] showing that prompt optimization can improve results generated by LLMs. As we move towards our vision for automated technical documentation translation powered by LLMs, other prompting strategies should be employed. III-C TriFid Preservation Scoring Framework Most ReadMe files in open-source software contain a combination of written instructions, code snippets, and URLs linking to additional or relevant resources. This information is typically formatted and structured using markdown [ 35 ] To assess how well Large Language Models (LLMs) preserve these key components, we introduce TriFid (TriFidelity) as an early stage evaluation framework. TriFid ’s evaluation and automated scoring is defined as follows: III-C 1 Code Block Evaluation and Score This score evaluates consistency in code blocks and their content after translation. Code Block Count Ratio: Let N_{\text{orig}} and N_{\text{trans}} be the number of code blocks in original and translated versions, respectively. The count ratio is: i. Case 1: If N_{\text{orig}}=0 and N_{\text{trans}}=0 , then r_{\text{count}}=1.0 ii. Case 2: Otherwise, r_{\text{count}}=\frac{\min(N_{\text{orig}},N_{\text{trans}})}{\max(N_{\text{orig}},N_{\text{trans}})} The code block count score is: CodeScore_{\text{count}}=50\times r_{\text{count}} Code Content Preservation: We verify that actual code content is preserved by normalizing code (ignoring translated comments and output lines starting with #, $, or *). Define the binary indicator: 1(\text{content preserved})=\begin{cases}1,&\text{if all normalized blocks in }\\ &\text{ original match translation}\\ 0,&\text{otherwise}\end{cases} The content preservation score is: CodeScore_{\text{content}}=50\times 1(\text{content preserved}) Combined Code Score: The overall code block score is: Score_{\text{code}}=CodeScore_{\text{count}}+CodeScore_{\text{content}} where Score_{\text{code}}\in[0,100] . III-C 2 URL Evaluation and Score This score evaluates URL preservation across original and translated texts, penalizing extra URLs in translation. Definition: i. U_{\text{orig}} = set of URLs in original text ii. U_{\text{trans}} = set of URLs in translated text iii. n_{\text{preserved}}=|U_{\text{orig}}\cap U_{\text{trans}}| iv. n_{\text{missing}}=|U_{\text{orig}}\setminus U_{\text{trans}}| v. n_{\text{extra}}=|U_{\text{trans}}\setminus U_{\text{orig}}| Total original URLs: N_{\text{origURL}}=n_{\text{preserved}}+n_{\text{missing}} Scoring Components: i. Preservation Component (70% weight): S_{\text{preserve}}=70\times\frac{n_{\text{preserved}}}{N_{\text{origURL}}} ii. Extra URL Penalty (up to 30 points): P_{\text{extra}}=\min\left(30,30\times\frac{n_{\text{extra}}}{N_{\text{origURL}}}\right) Bonus for no extra URLs: S_{\text{extra}}=30-P_{\text{extra}} Final Score: If N_{\text{origURL}}>0 : S_{\text{url}}=70\times\frac{n_{\text{preserved}}}{N_{\text{origURL}}}+\left(30-\min\left(30,30\times\frac{n_{\text{extra}}}{N_{\text{origURL}}}\right)\right) If N_{\text{origURL}}=0 : S_{\text{url}}=100 Result is rounded to nearest integer within [0,100]. III-C 3 Markdown Formatting Evaluation and Score This evaluation validates how well LLMs preserve Markdown syntax after translation [ 1 ] . Define M=7 categories: headers, bold, italic, lists, numbered lists, tables, and blockquotes. For each category i : i. E_{\text{orig}}(i) = count of elements in original text ii. E_{\text{trans}}(i) = count of elements in translated text For each category, the ratio is: i. If both counts are zero: r_{i}=1.0 ii. Otherwise: r_{i}=\frac{\min(E_{\text{orig}}(i),E_{\text{trans}}(i))}{\max(E_{\text{orig}}(i),E_{\text{trans}}(i))} The overall Markdown score is: S_{\text{md}}=\left(\frac{1}{M}\sum_{i=1}^{M}r_{i}\right)\times 100 IV Preliminary Findings IV-A Patterns of Translation Our findings indicate translation activities across the open-source software community. These translations spanned across multiple languages, demonstrating that global contributions are increasingly significant, as supported by GitHub’s recent report [ 11 ] . Our replication package [ 36 ] contains the extraction file showing the distribution of languages observed. We also observed some attempts to improve translation quality. In one notable case, an Italian speaker reported an issue indicating that the existing translation felt mechanical and lacked natural flow [ 37 ] . IV-A 1 Pull Requests After mining pull requests with “Translate ReadMe” included in the title, we found that for large-scale repositories (Section III-A ), there were 466 pull requests across 263 repositories with some repositories having more than one translation request. These pull requests included translations into multiple languages and updates to existing translations, reflecting a community-driven translation effort. For medium-scale repositories , we found 13 pull requests across 58 repositories, and for small-scale repositories , we found 24 pull requests across 500 repositories . Another relevant trend we observed with the timestamps of these pull request from 2014 to 2025 was that about 30% of all translation pull requests occurred in the month of October across large-scale repositories, which we believe may have been influenced by Hacktoberfest [ 25 ] , a month-long global open source contribution event. Figure [ 38 ] shows the patterns in translation activity trends throughout each year from 2014-2025. IV-A 2 Issues We also examined issues in repositories that included “Translate ReadMe” in their titles to see how frequently users and contributors raised related concerns. A similar trend emerged: In large-scale repositories , we found 64 issues across 260 repositories, in medium-scale repositories , 3 issues across 57 repositories; and in small-scale repositories , 4 issues across 500 repositories. Figure [ 39 ] shows the distribution of these issues. This results reveal disparities in translation activities across the life-cycle of open source repositories. Although translation efforts occur across levels of maturity, we observed a clear concentration of translation activities in large-scale repositories, though primarily opportunistic and likely influenced by project maturity and/or visibility. Table I shows the description of translation activities. TABLE I: Translation Activity Statistics by Repository Scale Repository Scale Repos PRs Issues Large Scale ( > 10k) 254 428 64 Medium Scale (5k-10k) 58 13 3 Early Stage (100-5k) 500 24 4 IV-A 3 Effectiveness & Sustainability of Translation PRs Across the large-scale projects we located 428 translation pull-requests. Most were accepted: 261 merged (61 %) vs. 167 closed. Table II summarises per-repository adoption. To understand post-merge upkeep we refined the sample to merged PRs that actually modified a README-like file that is, any file whose name matches README*.md or a language-tagged variant such as README.fr.md , README-zh_CN.md , etc. Table III shows, the English sources kept evolving (median 8.5 commits) whereas the corresponding translations received no follow-up commits, yielding a maintenance gap. TABLE II: Adoption (Merge Rate) among large repositories Metric Value Mean repo–level adoption rate 0.43 Median repo–level adoption rate 0.50 Repos with adoption \geq 0.80 18 Repos with adoption \leq 0.20 26 TABLE III: Six-month activity after merged translation PRs in Large Repositories (20 cases) English Translation Median commits 8.5 0 Mean commits 23.6 4.2 Maximum commits 166 0 For Medium-scale repositories we observed 13 PRs (11 merged); no repository showed a maintenance gap, that is, the English READMEs remained unchanged during the six-month activity window. Also, for Small-scale repositories, we observed 24 PRs (17 merged); mean adoption 0.65 , median 1.0 . Six translations lagged behind their English counterpart; four stayed in sync. IV-B LLM technologies for documentation translation Our evaluation of translations produced by ChatGPT 4.0 and Claude 3.5 Sonnet found that both models can process inputs or queries written in markdown and generate corresponding translations in markdown format when explicitly prompted. However, we observed notable differences in how accurately they preserve technical elements such as URLs, code snippets, and markdown formatting commonly found in technical documentation. To quantify this, we applied our early stage TriFid preservation scoring framework, which assesses the fidelity of translations in terms of code, URLs, and markdown structure compared to the original content. Based on the translation of 50 ReadMe files, Claude achieved an average accuracy of 95.6% across URL, Code, and Markdown Preservation scores, while ChatGPT 4.0 attained an average accuracy of 89.18% for the same categories. A closer inspection of markdown formatting revealed that 34% of translated ReadMe files exhibited inconsistencies in bold text, 32% lost italic emphasis, and 22% showed alterations in header structure. These discrepancies affect the integrity of the documentation’s information architecture, potentially hindering readability and navigability. We also observed a moderate to strong negative correlation between ReadMe file sizes and the models’ ability to preserve code, URLs, and markdown elements. URL preservation showed the strongest negative correlation at -0.699 , followed by markdown formatting at -0.499 . This indicates that as file sizes increased, the TriFid framework reported lower preservation scores across these dimensions. Table IV displays the average preservation scores across the 50 ReadMe files, while Table V provides detailed correlation values between file size and each TriFid preservation metric. TABLE IV: Average TriFid Preservation Scores Across 50 Readme translations Preservation Claude Avg. GPT Avg. Code Score 95.51% 85.24% URL Score 93.66% 89.52% Markdown Score 97.88% 92.80% TABLE V: Spearman Correlation Matrix between TriFid preservation scores and ReadMe file sizes Metrics Correlation Claude_Code -0.268958 Claude_Url -0.610617 Claude_Markdown -0.385345 GPT_Code -0.435587 GPT_Url -0.699682 GPT_Markdown -0.499684 V Discussion & Future Work We found that community engagement often plays a key role in supporting documentation translation within open source projects. However, our findings indicate that this support primarily benefits larger, more established repositories. Although periodic opportunities such as Hacktoberfest can trigger bursts of translation activity, smaller repositories often lack the visibility or contributor base needed for consistent translation efforts [ 40 ] . Even in more mature projects that do receive occasional contributions, sustaining and updating translations remains a significant challenge. LLMs present a promising, yet underexplored, alternative by offering on-demand translation capabilities that are independent of community size, event timing, or project popularity. While LLM-generated translations may not yet be fully reliable on their own, our results suggest they can serve as a strong foundation for a more systematic, AI-assisted approach to multilingual documentation. In cases where no translation currently exists, LLMs can produce initial drafts, easing the effort required to introduce diverse language support. For projects that update their original (typically English) documentation, change-detection tools could prompt LLMs to generate updated translations automatically. Moreover, as our findings indicate, providing LLMs with access to prior translations can enhance both quality and consistency, offering a path to improve translation updates over time. In both scenarios, communities can use our TriFid framework to evaluate and guide translation efforts thereby enabling current and potential contributors to maintain translation quality more effectively. V-A Fine-tuning LLMs for Documentation Translation Our early-stage TriFid evaluation uncovered structural issues that commonly affect LLM-generated translations of technical documentation particularly with preserving URLs, code snippets, and markdown formatting. Unlike general-purpose translation tasks, technical documentation requires attention to structural fidelity, accurate technical terminology, and completeness. As interest in LLMs grows, research increasingly supports the need for fine-tuning in domain-specific contexts [ 41 , 42 , 43 ] . A promising direction for improving translation of technical documentation lies in adapting pre-trained LLMs such as ChatGPT and Claude specifically for this purpose. This could involve designing loss functions that penalize structural deviations during training and incorporating verification mechanisms to ensure all sections from the source documentation are preserved in the output. V-B Optimizing TriFid While TriFid is still in an early stage without existing benchmarks for comparison, we plan to continuously optimize the framework going beyond ReadMe s to CONTRIBUTING guides and API reference pages. Since, TriFid provides crucial signals of preservation losses that are vital for fine-tuning LLMs for technical documentation tasks. The ability to easily refine and improve these core components makes LLM-driven translation more robust compared to traditional machine-translation approaches. V-C Human-In-Loop Even a carefully tuned TriFid cannot capture every nuance that makes a translation usable for developers. Adding a systematic human-in-the-loop layer on top of the automated checks. We will sample a small, stratified subset of PRs that pass (or narrowly fail) TriFid and send them to native speakers who are also professional technical writers or translators. These experts will (i) rate adequacy and fluency, (ii) flag subtler issues such as cultural tone, idiomatic terminology, or code comments that the metric currently overlooks, and (iii) annotate the minimal edits needed to reach publishable quality.