Armenian epigraphy, one of the world’s oldest and most diverse inscriptional traditions, remains largely absent from digital research infrastructures due to a lack of basic linguistic and conceptual resources. No machine-readable corpus, standardized terminology, or controlled vocabulary exists for describing Armenian inscription types, preventing indexing and interoperability. This paper addresses this gap by constructing the first dataset of Armenian inscription-type terminology and by developing a computational pipeline for analyzing it at scale. We digitize and preprocess a broad corpus of authoritative printed publications; curate a culturally grounded terminology list; and train transformer-based NER models to identify both attested inscription types and potential terminological variants across unseen texts. The resulting resources form the first empirical foundation for modelling Armenian epigraphic concepts needed for further developing a SKOS vocabulary aligned with, yet culturally distinct from, existing international epigraphic ontologies.

Armenian epigraphy is extensively documented in printed scholarly corpora, yet lacks machine-readable editions that support interoperability or computational analysis. In this paper, we present Armenian AutoEpiDoc, a system that automatically converts expert-verified Armenian inscription records into EpiDoc-compliant TEI/XML files. Operating on curated and domain-validated data, AutoEpiDoc maps Armenian-specific metadata to EpiDoc structures through rule-based templates and schema-aware validation. The workflow significantly reduces manual encoding effort and provides a scalable path toward producing digital editions and integrating Armenian inscriptions into international epigraphic infrastructures.

This study investigates how domain expertise shapes spontaneous oral language production, with a focus on architecture. Building on the ExpLay Corpus, which contains image descriptions by speakers with and without architectural training, we analyze linguistic variation by combining Profiling-UD and the DECAF framework. We extract a broad range of syntactic and morpho-syntactic features to build linguistic profiles for both groups and train classifiers to distinguish expert from non-expert productions. Two traditional machine learning models (logistic regression and SVM) are compared with a lightweight BiLSTM and two large language models (GliClass and LLaMA 2). While the expert and non-expert corpora diverge only subtly (pairwise Jensen–Shannon divergence (JSD)= 0.25), the BiLSTM using fastText embeddings achieves the highest F1-score (0.88), outperforming both traditional models and LLMs. This indicates that semantic representations are more predictive of domain variation than purely structural features and that smaller neural architectures generalize better on limited data. Overall, the findings provide empirical evidence that architectural expertise leaves measurable linguistic traces in spontaneous speech, supporting the Grammar of Space hypothesis.

We introduce CroCoSyn, a controlled, cross-lingual and cross-model corpus of 25,920 LLM-generated film synopses in English and French. Each synopsis is generated under systematically varied conditions, including model type, temperature, genre, protagonist gender, and narrative constraints, and enriched with structured metadata capturing characters and their relationships. Comparing Mistral and Llama across different model temperature degrees, CroCoSyn enables fine-grained analysis of narrative content, style, and character representation across models and languages. The corpus supports research on gender and cultural biases and story generation evaluation, and provides a foundation for comparative studies between LLM-generated and human-written narratives.

Despite the rapid growth of disinformation research, the fundamental reasons behind user engagement with such content remain poorly understood. Recently, several scholars have suggested that researchers should study engagement with disinformation as a form of collective action (CA). Drawing on Social IdentityTheory (SIT) and the Social Identity Model of Collective Action (SIMCA), this study empirically verifies this assumption by testing it across two distinct linguistic communities, English and Spanish. Specifically, it investigates whether mobilizing CA language functions as a uniform predictor of engagement, or if engagement is primarily driven by community specific identity dynamics. The experiment analysed a bilingual corpus of 4,035 X (formerly Twitter) posts associated with conspiracy theory and disinformation-related hashtags (e.g., #Agenda2030, #TheGreatReset). Using a mixed-methods approach combining BERTopic for narrative discovery, non-parametric statistical testing and Random Forest Regressor, we disentangled the effects of language presence from community behaviour. The results revealthat the Spanish community exhibits a higher baseline engagement compared to the English community indicating that engagement is primarily driven by macro-level community norms (i.e., identity) rather than micro-level linguistic triggers. We argue that rather than treating mobilizing language as a uniform predictor of engagement, future application of SIMCA in disinformation research should account for these identity-based baseline differences.

Named Entity Recognition has emerged as a critical task in natural language processing, particularly for extracting meaningful information from unstructured text. Although traditional approaches rely heavily on large annotated datasets, recent advances have explored weak supervision techniques to address the limitations of resource-intensive annotation processes. Historical texts provide unique challenges to this task because of their linguistic peculiarities, and several approaches exist to address texts of this domain in a supervised way, but they involve lengthy manual annotations of the documents of interest by domain experts. To address this issue, this paper explores how recent weakly supervised NER techniques can be adapted to historical texts, analyzing their suitability for this domain. The experiments show that domain-specific architectures can be effectively trained on low-resource corpora with weak supervision over a small set of entity labels. Using only 10% of the annotations, the performance of these architectures remains above 80% of the supervised quality in terms of F1-Score.

This paper investigates how language models (LMs) reproduce the existing gender disparity found in German media discourse about artificial intelligence (AI). Building on a human-annotated corpus of quotations from German media discourse on AI, we first quantify the frequency with which male and female speakers are directly cited across domains and speaker roles. We then train LLäMmlein (Pfister et al., 2025), a state-of-the-art German-only language model, GBERT, and a logistic regression model using only the quoted text as input and without providing any gender cues to classify the quotation as originating from a male or female speaker. By comparing model predictions with corpus-based gold labels, we find that male voices dominate both the corpus and the model predictions. Balancing the data mitigates but does not fully eliminate this disparity, indicating that the strong male-default tendency of transformer models cannot be explained by corpus skew alone, but also by their priors from pretraining. The study contributes to the interpretability of language models’ output for DH-related tasks, adaptation of NLP tools to domain-specific humanities corpora, and knowledge modelling in the humanities.

Linguistic diversity is increasingly under pressure globally and is becoming ever more relevant in digital contexts, where many languages remain structurally under-resourced, limiting access to language technologies and inhibiting equitable NLP development. To support linguistic diversity, publicly available data are needed that capture both the number of languages spoken and the distribution of speakers across them. We introduce GlobLingDiv, a database that uses country-level speaker distributions to derive language richness and entropy-based diversity measures, alongside a population-weighted digital language support measure. Applying these metrics globally, we examine the association between linguistic diversity and digital support conditions. The results reveal a substantial imbalance: highly diverse linguistic landscapes show comparatively low digital support, underscoring the need for more inclusive NLP environments.

Large Language Models (LLMs), even though exhibiting multiple capabilities on many NLP tasks, struggle with phonologically-grounded phenomena like rhyme detection and generation. When one moves to lower-resource languages such as Modern Greek, this is even more evident. In this paper, we present a hybrid neural-symbolic system that combines LLMs with deterministic phonological algorithms to achieve accurate rhyme identification and generation. We implement a comprehensive taxonomy of Greek rhyme types and employ an agentic generation pipeline with phonological verification. We use multiple prompting strategies (zero-shot, few-shot, Chain-of-Thought, and RAG-augmented) across several LLMs including Claude 3.7 and 4.5, GPT-4o, Gemini 2.0 and open-weight models like Llama 3.1 8B and 70B and Mistral Large. Results reveal a significant reasoning gap: while native-like models (Claude 3.7) perform intuitively (40\% accuracy in identification), reasoning-heavy models (Claude 4.5) achieve state-of-the-art performance (54\%) only when prompted with Chain-of-Thought. Most critically, pure LLM generation fails significantly (under 4\% valid poems), while our hybrid verification loop restores performance to 73.1\%. Along with the system presented, we further release a corpus of 40,000+ rhymes, derived from the \textit{Anemoskala} and \textit{Interwar Poetry} corpora, to support future research.

Authorship verification aims to assess whether a questioned text is stylistically compatible with an author’s known writings, a task that is particularly challenging in historical corpora with partial ground truth. We address this problem in the context of Mihai Eminescu’s journalistic corpus, a historically grounded collection comprising published articles, manuscripts, and texts of uncertain authorship. Using a profile-based framework with character n-grams and function words, we examine how stylistic compatibility behaves across different profile construction settings and temporal splits. The results show that character trigram profiles consistently accept verified texts while producing a small and stable set of rejections among disputed items, whereas function word profiles show near complete acceptance across the corpus. A qualitative analysis shows that rejected texts exhibit meaningful differences in discourse structure and communicative purpose. These findings illustrate how authorship verification can support literary scholarship through stable signals for close reading.

We study how to better use digitized historical archives to answer sociological and historical questions that require more context than raw text mentions provide. Using Finnish World War II Karelian evacuee family interviews, we build on prior extraction of 350K mentions of leisure activities and organizational memberships (71K unique names) that are too diverse and unstructured to analyze directly. We introduce a categorization framework capturing key dimensions of participation: type of activity/organization, typical sociality, regularity, and the level of physical demand. After creating a gold-standard annotated set, we evaluate whether large language models can apply the schema at scale and find that an open-weight LLM, combined with simple multi-run voting, closely matches expert judgments. We then label all 350K entities to produce a structured resource for downstream analyses of social integration and related outcomes.

Bibliographies are both humanities infrastructure and historic record. To computationally analyse them, however, requires implementing complex digitisation and standardisation decisions. This paper turns to Seyfettin Özege’s Eski Harflerle Basılmış Türkçe Eserler Kataloğu as an example, a scanned set of volumes marked by complex page layouts, degraded typography, irregular entry structures, and historically contingent inconsistencies. With this we present a pipeline that constructs a structured, machine-readable, and analysable dataset out of the 27,000 entries with computer vision, OCR, large and visual language models, sequence-based validation, and custom review tools. This process captures 97.8% of records, with remaining cases capable of being addressed by targeted review. This process demonstrates that combining LLMs with interpretable, review-centric pipelines, offers an appropriate approach for historically complex bibliographic sources.

Large language models (LLMs) are post-trained on human feedback collected from annotator communities, yet the linguistic influence of these annotator communities on language models remains poorly understood. We investigated the stylistic transfer from Nigerian annotators to the LLaMA family of models through a natural experiment with LLaMA 2 and LLaMA 3.1, as their release dates are separated by the shutdown of a major data annotation service provider in Nigeria. We generated corpora from both model families and measured linguistic style by computing the difference-in-difference of the Jensen-Shannon distance on the bigram distribution between model outputs and corpora of Nigerian English and US English. We found that, although both pre-trained model variants exhibit similar proximity to both English variants, the LLaMA 2 post-trained model moved toward Nigerian English, while the LLaMA 3.1 post-trained model moved away from Nigerian English. Qualitatively, we found that post-trained LLaMA 2 models used significantly fewer contractions, in line with Nigerian English speakers opting to use a formal register due to its role as an index of knowledgeability. Our findings suggest that annotator communities can imprint linguistic style on large language models, with potential implications such as a disproportionately higher false positive rate in AI plagiarism detection for users who share a linguistic style with annotator communities.

While context embeddings produced by LLMs can be used to estimate conceptual change, these representations are often not interpretable nor time-aware. Moreover, bias augmentation in historical data poses a non-trivial risk to researchers in the Digital Humanities. Hence, to model reliable concept trajectories in evolving scholarship, in this work we develop a framework that represents prototypical concepts through complex networks based on topics. Utilizing the Royal Society Corpus, we analyzed two competing theories from the Chemical Revolution (phlogiston vs. oxygen) as a case study to show that onomasiological change is linked to higher entropy and topological density, indicating increased diversity of ideas and connectivity effort.

Indirect speech is a fundamental yet understudied form of reported speech that plays a crucial role in literary texts and communication. While direct speech detection has received significant attention in computational linguistics, the automatic identification of indirect speech remains a challenge due to its nuanced linguistic structure and contextual dependencies. This paper focuses on the detection of indirect speech in late 19th-century Scandinavian literature, where its presence has been linked to shifting aesthetic ideals. We present an annotated dataset of 150 segments, each randomly selected from 150 different novels, designed to capture indirect speech in Danish and Norwegian literature. We evaluate four pre-trained language models for classifying indirect speech, with results showing that a Danish Foundation Model (DFM Large), trained on extensive Danish data, has the highest performance. Finally, we conduct a classifier-assisted quantitative corpus analysis and find that the prevalence of indirect speech exhibits fluctuations over time.

Automatic metaphor detection so far has largely focused on English data annotated for all kinds of metaphors including ubiquitous conventionalized ones. In this paper, we focus on deliberate metaphors in German sermons, i.e., metaphors that are used with a specific communicative goal. This task is harder because there is less training data available, and deliberate metaphors are very rare. Our goal is to support human annotators with automatically generated suggestions, so we strive above all for high recall. Using multilingual transfer learning based on various metaphor datasets and different transformer models, the highest recall we achieve is .70 (precision .10). Our results suggest that larger context windows beyond the sentence level are not helpful and that adding in-domain data even when annotated with different guidelines and in a different language is beneficial.

The present study introduces Semantic Factor Analysis (SFA), a novel computational approach recovering Big Five personality trait structures from pre-trained adjective word embeddings weighted by empirical participant data. Using Word2Vec embeddings trained on the Google-News-300 corpus, semantic relationships of IPIP-50 Big Five inventory adjectives (Goldberg, 1992) were extracted and factor structures computed through weighted vector averaging and K-means clustering. To validate the methodology, SFA was compared against a baseline using unweighted Word2Vec embeddings. In a controlled experiment with n=55 participants completing standard IPIP-50 assessments, HSP-R scale (Pluess et al., 2024) and multimedia impact surveys, empirically-weighted SFA successfully recovered all five personality dimensions with 62.5% average factor purity, substantially outperforming the unweighted baseline (52.0%, 10% relative improvement), while traditional Confirmatory Factor Analysis showed factor collapse and poor model fit. The approach was validated through Latent Class Analysis deriving empirically-based classification thresholds for Big Five dimensions and supporting a trichotomous Environmental Sensitivity model (Lionetti et al., 2018). Results demonstrate that integrating semantic representations with empirical data improves Big Five structure recovery beyond pure semantic similarity alone, particularly for small sample studies where traditional methods such as CFA will fail due to limited empirical data points.

While machine translation systems have been applied to many tasks with remarkable success, machine poetry translation has remained a challenge. This study investigates the capabilities of generative Large Language Models (LLMs) in the translation of poetry (taking Shakespeare’s 154 sonnets as an example) from English to German. For this purpose, I define metrics that assess the reproduction of the rhyme scheme and the metre of the original in a quantitative way. The results indicate that LLMs still lag behind professional human translators (especially with regard to the reproduction of the rhyme scheme), but that their performance is significantly influenced by the chosen prompt strategy. In particular, iteratively refining the result emerges as a successful strategy in terms of the reproduction of the form, but this comes at the expense of other aspects such as grammaticality and the reproduction of the meaning.

With the conflation of digital and non-digital spaces, and NLP technologies being integrated into an increasing number of aspects of daily life, linguistic diversity cannot be fully understood without considering how language is used online. While existing models of linguistic diversity typically have relied on speaker numbers or language production, the dimension of diversity in language consumption remains comparatively understudied. To facilitate such research, we introduce WikiLingDiv, an openly accessible dataset for quantifying linguistic diversity in online knowledge retrieval using Wikipedia page views. Our dataset is based on yearly page views of 340 language editions of Wikipedia, aggregated across 239 countries and territories over 10 years (2015-2024). Using the dataset, we illustrate spatial and temporal patterns of digital linguistic diversity, suggesting that diversity has both increased and decreased across countries and regions, while highlighting country-specific dynamics in language usage. We release the dataset as an openly available and easily integrable data resource for researchers in computational linguistics, digital humanities, and the broader social sciences, enabling further work on linguistic variation, digital inequality, and the interaction between language use and digital technology.

Although Wikipedia aspires to provide neutral information, alternative versions can be used for political manipulation. This paper analyzes how narratives about the Russo-Ukrainian War are linguistically reframed in a Russian Wikipedia Fork compared to the original Russian Wikipedia. Using Kullback-Leibler Divergence on a corpus of war-related edits in more than 13,000 articles, we identify key differences between the two versions. While the original Wikipedia features Ukrainian references and administrative details, direct war terminology, and Ukraine’s territorial designation, governance, and statehood, RWFork replaces or removes these elements, emphasizing reassignment of Ukrainian territories to Russia, favoring euphemistic war language, renaming locations, and recognizing Russia-backed DPR and LPR. These patterns closely align RWFork with demobilizational strategies observed in pro-Kremlin media.

The article focuses on a stylometric analysis of authentic literary texts and thematically related texts generated by large language models. The texts under study represent a fairly broad cross-section of twentieth-century French literature. Five models were used to generate the texts (ChatGPT 4-o, GPT 4-o mini, DeepSeek v.3, c4ai-command-r-plus, and c4ai-command-a). The original human-written stories of approximately 20,000 characters were summarized, and new narratives were then generated on the basis of these abstracts. In terms of plot and style, they were intended to resemble the originals. The research carried out with TF-IDF of the most frequent words showed that texts generated by specific LLMs and written by humans cluster relatively well as distinct groups. The experiments also showed that the "authorial" specificity of machine-generated texts partly matches the original clustering of human-written source texts.

Literary translation is rarely a neutral act of linguistic transfer, but rather a continuous series of conscious interventions - restructuring, semantic shifts, and stylistic adaptations. While Translation Studies analyzes these shifts qualitatively, current computational methods focus primarily on quality evaluation (e.g., BLEU, COMET) or authorship attribution (e.g., stylometry), lacking a scalable metric to quantify the extent and character of the translator’s intervention. We propose a novel method to measure the translator’s signal by using Interlinear Translation - a strict word-for-word gloss - as a computational baseline representing translational "Degree Zero," i.e., a neutral form of source text devoid of any stylistic adaptation.We define the Intervention Vector as the semantic difference between a literary translation and its interlinear counterpart in a high-dimensional vector space. We validate this approach on a multilingual corpus of the Greek New Testament translations comprising 5 interlinear baselines and 74 literary translations across 5 languages: English (16), French (14), Italian (12), Polish (16), and Spanish (16).Our results demonstrate that the magnitude of the Intervention Vector effectively ranks texts along a spectrum from literal to paraphrase, aligning with established theoretical categories. We find that this magnitude consistently distinguishes between translation strategies, yielding significantly longer vectors for dynamic and paraphrase strategies compared to literal and formal ones. This framework provides a quantitative method for analyzing translator agency without the need for a comprehensive corpus of reference translations.

Graph-based Retrieval-Augmented Generation (RAG) is increasingly used to explore long, heterogeneous, and weakly structured corpora, including historical archives. However, in such settings, naive full-corpus indexing is often computationally costly and sensitive to OCR noise, document redundancy, and topical dispersion. In this paper, we investigate corpus pre-targeting strategies as an intermediate layer to improve the efficiency and effectiveness of graph-based RAG for historical research.We evaluate a set of pre-targeting heuristics tailored to single-hop and multi-hop of historical questions on HistoriQA-ThirdRepublic, a French question-answering dataset derived from parliamentary debates and contemporary newspapers. Our results show that appropriate pre-targeting strategies can improve retrieval recall by 3–5% while reducing token consumption by 32–37% compared to full-corpus indexing, without degrading coverage of relevant documents.Beyond performance gains, this work highlights the importance of corpus-level optimization for applying RAG to large-scale historical collections, and provides practical insights for adapting graph-based RAG pipelines to the specific constraints of digitized archives.

This paper presents the first application of an automatic token-classification approach for detecting reported speech spans in Classical Latin using transformer-based neural architectures.Focusing on Seneca the Elder’s Declamatory Anthology, the study addresses the text’s highly polyphonic nature, resulting from theuse of reported speech. Instead of relying exclusively on sentence-level syntactic information, the proposed approach treats reported speech detection as a token-level sequence labeling problem. This enables the identification of reported speech spans extending across multiple sentences. We fine-tune three Latin neural language models —LatinBERT, LaBERTa, and PhilBERTa— for binary token-level classification and conduct experiments both with and without punctuation. The results show that RoBERTa-based models effectively identify reported speech, with LaBERTa achieving the best performance (F1 scores above 0.90).

We present the novel dataset GermAnProse, an annotated corpus consisting of four German short prose texts accompanied by an extensive set of narrative-focused annotations.As part of this dataset, we contribute an annotation scheme for mentions, speech, and character agency: Characters in Action (ChiA).GermAnProse also contains information on narrative phenomena: narrativity, semantic verb classes, and plot keyness.Moreover, we include reader reception data in the form of timing information for audiobook performances, indicating pauses between sentences and the time taken to read a specific sentence in a performance.We release the dataset, which contains more than 18,000 manually created standoff annotations in JSON format, enabling researchers to utilize this resource for further exploratory applications.

This paper investigates the lexicalisation of geographical nouns in Latin and Ancient Greek using a nd Ancient Greek using a diachronic, multi-genre corpus (8th cent. BCE – 2nd cent. CE) and Large Language Models for Word Sense Disambiguation. We focus on two main aspects: the onomasiological question of which words encode core geographical concepts, and the semasiological distribution of senses across lemmas. Across both languages, city-related concepts are the most frequently expressed, but Greek shows a stronger focus on maritime terms, whereas Latin favours concepts related to land. Semasiologically, Latin shows clearer evidence of semantic change over time (e.g., ’citizenship’ - ’city’, aequor ’flat surface’ - ’sea’), while Greek displays more gradual or distributed shifts. These results show that computational annotation enables cross-linguistic and diachronic analysis of spatial semantics, allowing us to compare the frequency of concepts across languages, genres, and periods, and to track when semantic change occurs and how core concepts evolve over time.

We propose a method to evaluate the extent to which an LLM’s observable input–output behavior aligns with established theories in the humanities and cultural studies. We instantiate the framework on three humanities theories—Davidson’s truth-conditional semantics, Lewis’s truth in fiction, and Iser’s concept of textual gaps—using a top-down, theory-driven black-box framework. Core assumptions of these theories are reconstructed into testable behavioral rules and assessed via controlled classification tasks with systematic prompt comparisons and significance testing. Our experiments show that theory-uninformed classification prompts generally outperform theory-enriched prompts in Lewis and Iser settings, while theory-informed prompts help in the Davidson task. Gemini Flash consistently achieves the highest scores across tasks and corpora, while the Iser gap detection task remains substantially harder than binary truth-conditional judgments. Statistical tests confirm robust prompt effects and the failure of basic prompts. However, model behavior under incremental theory exposure is unstable and architecture-dependent.

Although the context length of large language models (LLMs) has increased to millions of tokens, evaluating their effectiveness beyond needle-in-a-haystack approaches has proven difficult. We argue that novels provide a case study of subtle, complicated structure and long-range semantic dependencies often over 128k tokens in length. Existing novel-based long-context benchmarks are limited in scale due to the cost of manual annotating long texts. Inspired by work on computational novel analysis, we release the Too Long, Didn’t Model (TLDM) benchmark, which tests a model’s ability to reliably report plot summary, storyworld configuration, and elapsed narrative time. We find that none of seven tested frontier LLMs retain stable understanding beyond 64k tokens. Our results suggest language model developers must look beyond "lost in the middle” benchmarks when evaluating model performance in complex long context scenarios. To aid in further development we release the TLDM benchmark together with reference code and data.

We present an AI assistant designed to help researchers interact with language corpora using natural language instead of formal query languages. Built as a custom GPT with access to multilingual corpora via Czech National Corpus platform API, the system translates research questions into CQL queries, retrieves corpus data, and guides users through linguistic analysis. After more than a year of deployment, the system has processed over 1000 interactions with human users. We discuss the hybrid approach combining rule-based translation with LLM intelligence, challenges of building on a constantly evolving platform, and lessons learned from production usage. Notably, this system represents the first voice-enabled corpus interface in history, significantly lowering barriers to corpus-based research for non-technical users and users outside linguistic fields.

Biographical sources, such as literature encyclopedias, encode knowledge about historical figures in textual form. In this paper, we address the task of consolidating structured biographical information about authors from the former German Democratic Republic into a unified database. To this end, we present a generalizable Information Extraction (IE) system based on LLM prompting. Specifically, we compare two midsized open-source models, Qwen-2.5-32B and Llama-3-70B-Instruct, investigate a range of Prompt Engineering (PE) strategies, and propose a semantic similarity-based evaluation metric for open-ended IE. Our experiments on an unpublished annotated subset of biographical texts deliver moderate precision and variable recall, highlighting both the potential and current limitations of generative IE in the Digital Humanities.

This paper presents the design and construction of WikiFirst, a corpus for investigating the impact of content variation on authorship similarity under a fixed genre. Prior work has investigated individual authorial style and impact of genre. However, the role of content has remained underexplored due to the lack of suitable data. We address this gap by constructing a Wikipedia-based corpus consisting exclusively of first revisions authored by non-anonymous editors, thereby ensuring high authorship certainty while maintaining a stable encyclopaedic genre.

Prior studies on the symbolic power of languages have largely relied on surveys or localized experiments, limiting systematic comparison across cultures and domains. In this work, we propose an LLM-based multilingual persuasion simulation framework to quantify the symbolic power of languages through persuasion outcomes. We also introduce a Symbolic Power Index (SPI) that measures how language choice affects persuasion success and efficiency across domains. Experiments show that the LLM-based simulations largely reproduce established sociolinguistic prestige hierarchies tied to institutional authority and global power, especially in domains such as business, finance, education, and technology. These results suggest that LLM-based persuasion simulations offer a scalable, decision-making-driven approach to studying symbolic power in language.