linguistic typology

Natural language processing often treats linguistic variation as noise to be normalized, overlooking its sociolinguistic foundations and thereby compromising model robustness in real-world settings involving non-standard language forms. This work proposes the first systematic framework integrating sociolinguistic theory with NLP, advocating that linguistic variation—such as the orthographic diversity observed in Luxembourgish—should be modeled as a core feature rather than an artifact to be removed. The approach leverages sociolinguistically informed data construction, multivariate language modeling, and targeted fine-tuning to explicitly account for such variation. Experimental results demonstrate that ignoring linguistic variation significantly degrades model performance, whereas explicitly incorporating it enhances both generalization and robustness to non-standard linguistic forms.

This paper reveals that large language models (LLMs) as writing assistants induce systemic linguistic homogenization: they preserve semantic content while significantly reducing individual stylistic diversity, selectively amplifying dominant stylistic features and societal biases, and suppressing marginalized linguistic expressions. Employing a multimethod empirical approach—including controlled experiments, natural-text observation, quantitative stylistic analysis, bias classifier evaluation, and robustness testing across models, prompts, and scenarios—the study is the first to demonstrate the strong generalizability of this phenomenon. Key contributions are: (1) establishing that LLM-driven linguistic diversity erosion poses profound risks to fairness (e.g., misjudging cultural adaptability in hiring), clinical diagnostics (loss of individuating language cues), and cultural preservation; and (2) providing the first reproducible, multidimensionally validated framework for assessing the sociolinguistic impact of AI-mediated language intervention.

This study addresses the persistent performance disparities of multilingual language models across languages, investigating whether these gaps stem from inherent linguistic complexity or modeling design choices. For the first time, it systematically disentangles these factors by jointly analyzing modeling mechanisms—including tokenization, encoding strategies, data sampling, and parameter sharing—alongside linguistic properties such as morphology, syntax, and information density. The findings reveal that the majority of cross-lingual performance gaps are attributable to modeling decisions rather than intrinsic language characteristics. Building on this insight, the work proposes actionable principles for fairer model design and demonstrates that standardized tokenization, unified encoding, and balanced data exposure significantly enhance linguistic equity in multilingual systems.

This study investigates whether four morphosyntactic variants associated with pronouns in Brazilian Portuguese can reliably indicate speakers’ regional dialect origins. Integrating sociolinguistic theory with computational methods, the research develops a model of morphosyntactic covariation and systematically evaluates the effectiveness of correlation analysis versus clustering algorithms in uncovering regional dialect patterns. Findings demonstrate that clustering methods more effectively capture geographically coherent groupings, significantly outperforming traditional correlation-based approaches. The work not only confirms the diagnostic value of morphosyntactic covariation for inferring regional provenance but also offers a scalable, interdisciplinary analytical framework for dialect identification, thereby advancing the integration of computational linguistics and sociolinguistics.

Low-resource languages encode vital cultural and historical knowledge yet suffer from data scarcity, inadequate model adaptation, and insufficient cultural sensitivity. To address these challenges, we propose the first large language model (LLM) application framework tailored for humanities research on low-resource languages. Our method integrates instruction fine-tuning, few-shot prompting, multilingual knowledge distillation, and cultural-context alignment, augmented by domain-specific knowledge graphs and sparse-label enhancement to enable culturally grounded fine-tuning and ethics-aware data governance. Experimental results demonstrate that our customized models achieve 32–57% accuracy improvements over baselines on three core digital humanities tasks: classical text transcription, endangered dialect analysis, and oral history structuring. As a community resource, we release LinguaHumanis v1.0—an open-source, task-diverse evaluation benchmark—providing both methodological foundations and practical implementation guidelines for low-resource language research in the digital humanities.

This study addresses key challenges faced by social science researchers when using large language models (LLMs) for text annotation—namely, poor reproducibility, annotation errors that compromise statistical inference, and high technical barriers. To overcome these issues, the authors propose the first end-to-end LLM-based text annotation framework tailored specifically for the social sciences and humanities (SSH). The framework integrates structured prompt engineering, open-source LLM API integration, cross-validation, and error propagation modeling, with an explicit emphasis on avoiding prompt overfitting and quantifying annotation uncertainty. Implemented in both Python and R, this approach establishes a transparent, reproducible, and scalable workflow that substantially enhances the reliability, efficiency, and methodological rigor of automated text annotation in SSH research.

Current large language models (LLMs) are applied in linguistics in a fragmented, ad hoc manner, lacking systematic methodology and theoretical integration. To address this, we propose the first two-tiered, linguistics-oriented methodology framework, comprising three goal-aligned paradigms: prompt-based interaction, fine-tuned modeling, and embedding probing. The framework integrates prompt engineering, open-weight model fine-tuning (e.g., LLaMA), context-aware embedding quantification, and multi-stage research pipeline design. It prioritizes reproducibility, verifiability, and theory-driven inquiry. Empirical validation employs retrospective analysis, prospective experimentation, and expert surveys. Results demonstrate substantial improvements in methodological reproducibility and theoretical rigor, enabling linguistics to transition from empirical LLM application toward a robust, scientific paradigm grounded in computational and cognitive linguistic principles.

This study addresses the longstanding limitations in Arabic natural language processing (NLP), which stem not from linguistic complexity per se but from insufficient engagement with sociolinguistic realities and a lack of interdisciplinary integration. Through a systematic review of two decades of research, the work identifies critical gaps in infrastructure and paradigms, attributing them to inadequate social embedding and cross-disciplinary collaboration. By synthesizing efforts in language resource development, shared task organization, social media analysis, and computational social science, the project elucidates the challenges of transfer between Modern Standard Arabic and its dialects. It further underscores the sociocultural attributes of datasets and the pivotal role of task-oriented communities. The study advocates for a paradigm shift in low-resource NLP—one that transcends purely technical solutions to holistically address social, institutional, and cognitive dimensions—offering a new framework for global low-resource language research.

This study addresses the current lack of interdisciplinary understanding regarding the integration pathways, efficacy boundaries, and systemic risks of large language models (LLMs) across natural sciences, social sciences, and humanities. Through a systematic literature review and illustrative case analyses, it critically evaluates the deployment of LLMs throughout the research lifecycle—including hypothesis generation, literature synthesis, data analysis, and scholarly writing. The work identifies ten previously underappreciated systemic risks, such as diminished researcher autonomy, AI-induced confirmation bias, ambiguous authorship, and inequitable access to technology. It further demonstrates how LLMs, while enhancing efficiency, simultaneously introduce challenges like hallucination, irreproducibility, data bias, and model opacity. To guide responsible adoption, the study proposes an interdisciplinary governance framework and a roadmap for explainable AI research in scholarly contexts.

This study addresses the limited understanding of the intrinsic structural properties of large language models (LLMs) in multilingual processing, particularly the systematic differences between low-resource and high-resource languages such as English. Moving beyond prior work that primarily focuses on token-level representations, this paper pioneers a language-structure-oriented perspective by employing representational structural analysis combined with cross-lingual representation comparison and structural similarity metrics. The findings reveal that low-resource languages exhibit significantly divergent internal structures compared to English within LLMs, and that the degree of structural similarity strongly correlates with language resource availability. Furthermore, language-specific post-training is shown to effectively reshape internal representations while preserving inter-language relationships, thereby uncovering the formative role of post-training in shaping the multilingual structural geometry of LLMs.