Modeling communicative intentions in therapist–client dialogue using large language models and graph analysis

We introduce a coarse-grained, turn-level, intention-centered computational framework for large-scale analysis of communicative intentions in psychotherapist–client dialogue, integrating intentional coding, automated intention classification using open-source large language models, and graph-based structural modeling. The study is based on a formal Intentional Coding System of Psychotherapeutic Discourse (ICSPD) and an annotated corpus of psychotherapeutic dialogues, enabling large-scale analysis of therapist–client interaction patterns. Open-source language models were employed for automated annotation, demonstrating robust performance in intention classification, thereby supporting transparent and accessible research workflows that do not rely on proprietary APIs. Among openly-licensed models (Apache 2.0/MIT), the fine-tuned Qwen3-8B model achieved the highest performance, with weighted F1-scores of 0.80 for therapist intentions and 0.85 for client intentions. The model excelled at identifying information-exchange intentions but exhibited reduced accuracy for emotional and evaluative categories. Graph-based modeling of multi-turn interaction patterns across a corpus of 30,724 therapist–client utterances, annotated using fine-tuned intention-classification models trained on an expert-labeled subset (4325 utterances), revealed a structured network of communicative intentions characterized by recurrent communities and transition patterns. Community detection using the Louvain algorithm identified four recurrent interaction clusters that were descriptively interpreted as supportive, reflective, interpretative, and problem-solving configurations of therapeutic communication. Notably, the reflective–exploratory cluster emerged as a central hub linking supportive, interpretative, and problem-solving interaction configurations. Overall, the proposed approach provides a computational framework for investigating psychotherapeutic interaction at the turn-level and demonstrates that open-source language models can support transparent and explainable automated analysis of therapeutic discourse for corpus-level research and training applications when integrated with graph-based methods.