Linguistic markers of professional burnout in employees' self-descriptive texts: a comparative corpus analysis and the limits of applicability of dictionary-based AI methods

This study presents a comparative corpus analysis of 37 English-language self-descriptive texts written by employees (7,053 tokens): Group A (burnout, n = 20) and Group B (engagement and satisfaction, n = 17). Hypotheses derived from a preliminary AI-generated analysis were tested at the lexical, syntactic, orthographic/punctuation (including emojis) and pragmatic levels, using rates per 1,000 tokens, Dunning’s log-likelihood ratio and text-level Fisher’s exact tests. The groups differed in positive-affect vocabulary (5.2 vs. 16.8‰), exclamation marks (1.7 vs. 11.5‰), pictographic markers (0.2 vs. 3.7‰), obligation modality (3.9 vs. 0.4‰), ‘can’t/cannot’ constructions (2.2 vs. 0‰) and absolutist words (10.4 vs. 4.1‰). First-person singular density (92.7 vs. 95.3‰) does not distinguish between the groups, and the putative burnout markers “!”, ALL CAPS and the emojis 🙁 / 😢 were not confirmed. An analysis of the limitations of dictionary-based tools (such as LIWC) — irony, scare quotes, ambivalence and orthographic noise —justifies the development of a GDPR-compliant, anonymous, aggregate-only HR dashboard.