Knowledge Discovery of Emerging Developer Skill Profiles in Telegram Community Using Semantic Embeddings

Abstract

This article proposes a pipeline to discover professional profiles of software developers from job posts published on Telegram. The corpus consists of 22,243 messages collected between May 2021 and January 2026 from a public Brazilian group dedicated to IT vacancies; a two-stage filter combining deterministic rules and semantic validation reduced the set to 7,264 valid posts. A taxonomy of 108 technical and contractual competences was applied to the corpus, three embedding models were compared, and the reduced space was partitioned with K-Means. The number of clusters followed internal indices restricted by admissibility criteria defined beforehand and decided by an independent evaluator, resulting in 23 profiles with mean topic coherence of 0.522. Labelling and auditing relied on models from different families, so that no model judged its own output. This choice mattered: the cross-architecture protocol spread the scores instead of concentrating them, and showed that about a fifth of the corpus forms clusters built around posting format rather than occupational content. With each competence expressed as a share of monthly posts and multiple-testing correction applied, only three of 50 series decline significantly and none rises.