Graph-context model for semantic representation of dialogues in digital psychological support services

Abstract

The article addresses the problem of semantic representation of dialogues in digital services for psychological support of the population. The relevance of the study is determined by the fact that messages in such services are often short, incomplete, and context-dependent, while their meaning is defined not only by the text itself, but also by the author’s role, previous utterances, thematic features, emotional markers, and the current dialogue state. The aim of the study is to develop a graph-context model for semantic representation of dialogues that provides a formalized description of utterances, participant roles, thematic features, emotional markers, communicative intentions, dialogue states, and the relationships between them. The proposed model represents a dialogue not as a linear sequence of isolated messages, but as a heterogeneous semantic structure. Within the model, the input structure of a dialogue, the semantic profile of an utterance, sets of nodes and edges, edge types, a mechanism for weighting contextual dependencies, and the representation of a dialogue as a trajectory of semantic states are formalized. The semantic profile of an utterance combines a textual representation, the author’s role, the utterance position in the dialogue, local context, thematic and emotional components, communicative intention, uncertainty level, and a service feature. The rules for typing graph nodes and edges are separately defined, which makes it possible to distinguish sequential, role-based, thematic, emotional, intentional, state-related, and contextual dependencies between dialogue elements. The proposed representation can be used as an input formalized object for further utterance classification, dialogue state identification, analysis of transitions between states, and construction of an explanatory substructure of the result. The experimental study showed that the use of the graph-context representation improves the quality of utterance classification and dialogue state identification compared with textual and linear-context approaches. The Macro-F1 score for utterance classification and the TransitionAcc score for identifying transitions between dialogue states were both zero point seven eight. The obtained results confirm the feasibility of using the graph-context model as a basis for automated semantic analysis of dialogues in digital psychological support services.