Corpus Linguistic Methodology with Sketch Engine

• Corpus Linguistic Methodology is a systematic approach that uses computational tools to analyze large text corpora through statistical, collocational, and grammatical techniques.
• It emphasizes robust corpus construction and automated preprocessing, enabling precise extraction of keyness and collocational profiles from diverse text sources.
• Insights from this methodology, exemplified by Sketch Engine, inform both technical design in VR applications and nuanced clinical discourse analysis.

Sketch Engine is a corpus query and linguistic analysis platform widely utilized in corpus linguistics to analyze large text corpora with advanced statistical, collocational, and keyness-based metrics. In research analyzing online discourse about virtual reality (VR) and anxiety, Sketch Engine was employed to extract high-frequency words, collocates, and keyness profiles, revealing dominant conceptual and technical patterns associated with VR-based anxiety discussions (Yamoah et al., 7 Dec 2025). The following sections detail its technical principles, methodologies, corpus construction strategies, analytical metrics, applied findings, and implications for technical research domains.

1. Technical Foundations and Core Functionality

Sketch Engine operates as a web-hosted corpus management suite supporting multi-billion-token corpora across numerous languages. Its distinguishing features for the arXiv audience include:

• Industrial-scale tokenization, lemmatization via tools such as MorphoDiTa, and part-of-speech (POS) tagging.
• Robust query interfaces for word frequency, keyness, collocation profiles, grammatical sketches ("Word Sketch"), and concordancing.
• Support for custom corpus uploads, dynamic subcorpus extraction, and query scripting.

Sketch Engine’s analytical engine is optimized for both research-scale corpora (e.g., English Trends, 86 billion tokens) and project-specific subcorpora (e.g., VR–anxiety discourse subsets).

2. Corpus Construction and Preprocessing

In applications such as the study of VR and anxiety online discourse (Yamoah et al., 7 Dec 2025), Sketch Engine’s workflow proceeded as:

• Automated filtering to select documents containing both VR-related keywords (“virtual reality”, “Oculus”, “headset”, etc.) and “anxiety”.
• Extraction from the “English Trends” monitor corpus, aggregating news, popular media, and Wikipedia from 2014–2025.
• No additional preprocessing beyond native tokenization, lemmatization, and stop-word marking, as Sketch Engine applies these by default on ingest.

Custom subcorpora derived in this manner routinely exceed tens of millions of tokens (e.g., VR–anxiety subcorpus ∼34.7 million tokens).

3. Key Statistical and Collocational Metrics

Sketch Engine incorporates multiple statistical routines foundational to large-scale corpus analysis:

A. Keyness

Keyness quantifies the over- or under-representation of a word (lemma) in a focus corpus (F) versus a reference corpus (R). Yamoah & Dykeman (Yamoah et al., 7 Dec 2025) apply Kilgarriff’s “Simple Maths” keyness, defined as:

$$\text{Keyness}(w) = \left( \frac{f_F(w)}{N_F} - \frac{f_R(w)}{N_R} \right) \times 10^6$$

where $f_F(w)$ and $f_R(w)$ are raw frequencies in F and R; $N_F, N_R$ are corpus sizes.

Positive keyness implicates topical salience within the subcorpus; negative values denote under-representation.

B. Collocation Analysis and logDice

Collocate analysis profiles word associations in defined windows (±5 tokens). Strength of association is quantified by logDice—preferred for its stability across corpus sizes and interpretability in log scale:

$$\text{logDice}(w, x) = 14 + \log_2 \left( \frac{2f(w, x)}{f(w) + f(x)} \right)$$

where $f(w, x)$ equals the joint frequency, and $f(w), f(x)$ are marginal frequencies. logDice ranges from near 0 (random association) up to 14 (complete overlap).

Sketch Engine also supports PMI, log-likelihood (G²), and t-score, though these were not explicitly reported in (Yamoah et al., 7 Dec 2025).

C. Concordancing and Grammatical Sketches

Word Sketch provides automatic extraction of grammatical relations (modifiers, objects, prepositional phrases) allowing multi-dimensional portraiting of a lemma’s usage.

4. Application: Analysis of VR–Anxiety Online Discourse

Yamoah & Dykeman (Yamoah et al., 7 Dec 2025) use Sketch Engine to produce empirical results regarding VR and anxiety language networks:

• High-keyness terms predominantly reflect hardware and immersive apparatus (“VR”, “Oculus”, “headset”, “Vive”, “AR”), with “anxiety” ranking below major device names.
• Collocational clusters for “virtual reality” uniformly point toward technical, experiential, and design-centered usage (e.g., “immersive”, “location-based”; “of/in/for virtual reality”).
• Medical and technical phraseology surrounding “anxiety” reflects clinical discourse, with frequent compounds like “anxiety disorder”, “anxiety reduction”, “generalized anxiety”, highlighting the diagnostic and outcome focus.

Table: Top Keyness Lemmas (2014–2025 VR–Anxiety Subcorpus vs. Reference)

Rank	Lemma	Keyness
1	VR	931,354
2	Oculus	47,099
3	headset	105,294
4	Vive	17,397
12	anxiety	51,484

Only values reported in (Yamoah et al., 7 Dec 2025).

Such quantitative patterns substantiate that VR–anxiety discourse is technically biased toward device specification and immersive attributes, with clinical keywords embedded.

5. Significance and Research Implications

Sketch Engine enables advanced, reproducible corpus-based investigations into both technical and clinical discourse. In VR–anxiety studies (Yamoah et al., 7 Dec 2025), the methodology affords:

• Precise mapping of dominant device-related terminology and collocational themes.
• Quantitative comparison of experiential versus diagnostic language in technical communities.
• Modelling of discourse evolution via keyness and collocation dynamics across years, hardware generations, and thematic shifts.
• Informing design guidelines for in-VR therapies reflecting real-world user concerns (e.g., hardware comfort, session parameters suggested by corpus frequency of “headset weight/discomfort”).

6. Limitations and Directions for Advanced Corpus Linguistics

While robust, Sketch Engine’s approach exhibits several constraints (Yamoah et al., 7 Dec 2025):

• Register bias is prominent; corpora drawn from news/media emphasize technical apparatus over clinical outcomes.
• Subcorpora based on document-level keyword filtering may omit granular therapy discussions in patient forums or clinical literature.
• logDice and keyness only capture direct lexical associations, not broader semantic or pragmatic networks.

Future research should:

• Integrate supplementary corpora from clinical forums, practitioner reports, and patient testimonials.
• Apply additional association metrics (PMI, log-likelihood) to uncover lower-frequency but therapeutically salient collocations.
• Drive cross-disciplinary annotation (hardware taxonomy, intervention protocols) to further disaggregate technical versus clinical VR–anxiety language.

7. Role in Technical and Applied Linguistics Research

Sketch Engine remains the reference linguistic tool for rapid, large-scale corpus query and analysis. Its capacity to quantify keyness, model collocational structure, and automate grammatical relation mapping supports the construction of domain-specific subcorpora for both technical (device-focused) and clinical (therapy-focused) applications. This functionality is foundational for corpus-driven text mining, lexicography, and natural language processing workflows in the VR–anxiety research domain and beyond (Yamoah et al., 7 Dec 2025).

---

In summary, Sketch Engine serves as the primary analytical platform for quantitative corpus-based linguistic analysis, furnishing robust statistical metrics, collocational profiles, and comprehensive grammatical sketches. Its application to VR and anxiety discourse yields data-driven insight into technology-centered and medical linguistic ecosystems, providing actionable intelligence for hypothesis-driven research and system design.