This study investigated the utility of lexical sophistication indices for proficiency detection. Features were drawn from two distinct accounts of lexical organization: repetition-based and context-based. Learner data from the EFCAMDAT corpus, subsampled for topic balance, were classified across adjacent CEFR levels using elastic net regularized logistic regression. Modeled separately, word frequency indices, featuring broader corpus coverage, partially outperformed the more computationally diverse context- and semantics-based indices. Modeled together, performance exceeded that of separate models and approached or improved on previously reported proficiency detection performance. Important features in the model largely represented CD/SD indices, contextual redundancy and lexical strength, while indices more commonly relied on in lexical sophistication studies notably underperformed. This study sheds light on a theory-informed feature selection approach while highlighting critically unexplored indices and corpus sources (e.g., social media) for future studies and applications of lexical sophistication and proficiency detection.
This article reports on a corpus-based study examining vocabulary of textbooks used as reading materials in twelve subjects under the domain of mechanical engineering at a higher education institution in the China context. The present study aims to exam the vocabulary demand of mechanical engineering textbooks and to create a Mechanical Engineering Word List to cover the EFL undergraduates’ lexical shortage. It first presents the vocabulary profiles of subject texts from the domain of mechanical engineering using Nation’s (2020) British National Corpus/Corpus of Contemporary American English (BNC/COCA) word frequency lists. Then, it outlines the vocabulary load required to comprehend different subjects of texts within twelve sub-corpora. The results indicate that students need to contain at least 4th to 7th BNC/COCA list word family to reach the 95% lexical comprehension, which is too much burden. The findings suggest that lexical demands vary when handling reading materials across different subjects in mechanical engineering, and provide insights into the extent of vocabulary knowledge needed to understand and learn different subject content through texts. After analysis of the lexical load of textbook texts across subjects, a Mechanical Engineering Word List (MEWL) is needed to help students acquire the 95% lexical coverage to reach the minimally adequate comprehension with less reading burden.
Corpus Sense is a web application for content and discourse analysis designed to make advanced corpus exploration accessible to a broad range of user profiles, from corpus linguists to social scientists with no technical background. The original version, described in Moreno-Ortiz (2025), combined traditional corpus analysis methods with natural language processing and AI-assisted functionalities: keyword extraction, named entity recognition, semantic search, topic modelling with labels generated by a large language model (LLM), and LLM-generated insights. This paper presents version 2.0, a major release that re-engineers the platform and extends its analytical tool set. The monolithic implementation of the first version has been replaced by a modular platform with asynchronous background processing: analytics become available as soon as linguistic parsing completes, long-running analyses no longer tie the user to an open browser tab, and new tools can be added without structural rework. On this foundation, version 2.0 delivers the functionalities identified as priorities in the original paper: grammatical collocation analysis with an interactive navigator, n-gram extraction with C-value scoring and interactive curation, transformer-based sentiment and emotion classification with corpus trajectories, and metadata-based subcorpora, together with corpus-wide dispersion and evenness statistics and a redesigned WordMap visualisation. Every tool from the first version has been revised. The paper describes the new architecture, the revised and new tools, and the platform’s AI-assisted features; it situates the collocation and n-gram tools against Sketch Engine, AntConc, WordSmith Tools, CQPweb and LancsBox, including an empirical comparison, and closes with a roadmap of further modules.
This study applies corpus linguistic and machine learning methods to analyze 300 Finnish children’s books intended to be read between the ages of 7 and 15, and examines the linguistic characteristics and distinctiveness of texts written for particular ages. The corpus we use, the Turku Children’s Book Lexicon (TCBC), consists of three sub-corpora based on the intended reading ages of the books (7-8, 9-12, 13-15) and we use two different methodologies to analyze the linguistic characteristics of texts in these age groups. First, we use Key Feature Analysis (Egbert and Biber 2023) to examine the typical functional features of texts on the corpus level. Second, for examining the distinctiveness of the text classes on the level of individual texts, we use interpretable machine learning methods and Support Vector Machines (SVM) in particular. In our analysis, we use a total of 148 features. We show that the texts in each age group have distinct linguistic characteristics, with differences in phrasal and clausal complexity, morpho-lexical richness, and lexical variation. However, classifying individual texts proves quite difficult and machine learning models struggle especially with distinguishing texts of the middle age group from the others. We also show empirical evidence that traditionally used readability metrics are insufficient for characterizing text classes. This study provides crucial information for a wide range of fields applying age-specific text data, such as psycholinguistic reading studies and the development of education technologies.
This systematic review synthesizes empirical studies (n=27) on the integration of corpus-driven and AI-enhanced language pedagogies (CD-AIE-LPs) in language learning. Following the PRISMA 2020 guidelines, the review addresses four key issues involved in CD-AIE-LPs: the instructional foci, technological implementations, theoretical and methodological foundations, and pedagogical effectiveness. It uncovers four main findings: (1) although a range of language features has been examined, instructional targets focus predominantly on vocabulary, grammar, and academic writing, with limited attention to speaking, listening, and pragmatic competence; (2) widely used corpus tools (e.g., COCA, AntConc) and generative AI models (e.g., ChatGPT, Gemini) constitute the core technological infrastructure, with loose-sequential integration emerging as the dominant model, while retrieval-augmented generation (RAG) represents an emerging deeper integration model; (3) research relies heavily on quasi-experimental and mixed-methods approaches; and (4) overall, CD-AIE-LPs are effective in enhancing learning outcomes, with the combined use of corpus and AI tools consistently outperforming either technology used alone. The review concludes by discussing pedagogical and research implications of the findings and key limitations of the current research, including inadequate theoretical grounding, relatively small and homogeneous samples, and a scarcity of deeply integrated AI–corpus systems, thereby highlighting important priorities for future research.
