GEMCo: A Validated, Ethically Releasable Proxy for Inaccessible Counselling Data

Computational research in mental health, such as developing triage bots or training support systems, is severely hindered by the lack of accessible, high-quality dialogue data. Because real counselling sessions are protected by strict privacy and ethical regulations, researchers often lack the datasets necessary to build or evaluate tools for professional practice. While English-language resources exist, there has been a notable absence of natively German, multi-turn, asynchronous counselling corpora that can be shared openly with the research community.

To address this gap, the authors developed GEMCo, a dataset consisting of 86 complete, human-written e-mail counselling threads. The corpus is divided into two parts: GEMCo-A, which contains 50 expert-authored cases created by experienced counsellors, and GEMCo-B, which comprises 36 threads generated through role-play sessions between professional counsellors and trained students. By using human-authored content rather than machine-generated text, the authors ensure the data maintains the nuance and structure of professional practice while remaining entirely free of real personal information, thus allowing for a CC BY 4.0 release.

To ensure GEMCo is a reliable stand-in for real data, the authors validated it against a held-out reference set of 124 authentic, anonymized counselling conversations. They employed a rigorous pipeline that segments text into semantic spans and classifies them based on counsellor strategies (using the OnCoCo taxonomy) and client emotions (using Ekman’s categories). The validation compares the proxy to the real data using the Jensen–Shannon divergence (JSD), scaling the results against the natural variation found within the real data itself (split-half reliability). This approach allows researchers to determine if the proxy's deviations are within the range of natural noise or represent a significant departure from real-world practice.