Unveiling Power Dynamics in Relational Recovery: A Computational Analysis of Therapeutic Dyad Interviews

Recovery from severe mental distress emerges through relational processes co-constructed between individuals and their support networks. A central aspect of these processes concerns latent power dynamics within the relationships aimed at recovery. While research on these dynamics has been predominantly qualitative, the field lacks quantitative methods to measure how power asymmetries manifest linguistically. This study develops a multilayered computational framework to measure and represent relational power dynamics in mental health narratives. We apply a hybrid approach combining corpus analysis with LLM-based methods to interviews collected within the EXIT MADNESS project; participants (utenti) in stable recovery (5+ years) were interviewed with facilitators (facilitatori) they identified as most significant to their recovery, both in-pairs and individually. First, a descriptive corpus analysis through basic statistics, correspondence analysis and keyness ananlysis was conducted on all interviews (25 dual, 50 individual) in order to inspect some key patterns within the corpora. Then, the analysis of power dynamics was conducted exclusively on the dyadic interviews, by integrating lexical-surface indices, with large language model (LLM-as-a-judge) evaluations. Results show that dyadic interviews occupy an intermediate yet distinct lexical space, supporting the idea of co-constructed discourse. Both indices and LLM evaluations capture power as dynamic and temporally evolving consistent with our theory. Overall, the study bridges qualitative and computational approaches, and suggests AI-based tools for analyzing relational dynamics in therapeutic settings.