The paper is dedicated to crucial tasks of communication analysis in high-load contact centers, including structured data collection, key events identification, and automation of speech pattern analysis. Emphasis is placed on the optimization of dialog scripts through topic-based graph modeling to discover the most effective pathways for agent-customer communication, as well as dynamic classification of long multi-intent conversations to identify subtasks and improve the accuracy of complex communication analysis. The authors suggest a methodology that combines graph algorithms for identifying standard scenarios and extracting anomalies in dialog corpora, as well as n-gram segmentation of dialogs with adaptive overlapping for context preservation and dynamic classification by using lightweight Large Language Models and aggregation rules. The developed toolkit was tested on real-life technical support chat and call data. Results have shown its effectiveness for monitoring and analyzing agent-customer interactions, allowing for up to 100% coverage of dialogs analyzed. Practical significance of the study is attested by use cases such as common speech pattern extraction, event-topic context analysis, and retrieval of similar dialogs based on visual representations of task type dynamics.