German Counseling Grounding-Act Corpus (GRACO)

Abstract

We present a corpus of 196 German counseling conversations (ca. 25k turns) between advice seekers and counselors from nine domains. A subset of 11.5k turns was double-annotated with grounding acts (e.g., acknowledgments, repairs), attempts to advance the conversation, success of advancing, and conversation phases. Baseline classification experiments with logistic regression and GBERT-base illustrate the impact of class imbalance in grounding-act classification. For logistic regression, train-only balancing improves Macro-F1 from 0.417 [0.377–0.434] to 0.444 [0.394–0.478]. For GBERT-base, performance remains competitive (Macro-F1 0.481), with balancing yielding comparable results under the same evaluation protocol. Given the scarcity of German corpora of naturally occurring conversations annotated for grounding phenomena, we provide a novel resource for both conversation analysis and natural language processing, facilitating the design of realistic human-language model interactions in German. Code and data are available at https://osf.io/6k275/overview.

Resources

Details