A Multiscale Computational Framework for Structural Organization in Long-Form Chinese Podcast Interviews

Long-form podcast interviews require methods that distinguish episodes, speaker streams, windows, and chronological turns. This study presents a multiscale computational framework for 53 episodes of the Chinese podcast Yan Zhong Hua Shu (11,590 turns) and 40 external episodes (20 dialogic, 20 monologic) from four podcasts. Methods combined dependency parsing, Shannon entropy, fixed-token resampling, topic modeling, embeddings, and turn-level lexical similarity. Both roles showed negative mean entropy slopes, while the average host trajectory declined more markedly near the beginning. Negative slopes persisted under 384-token control and alternative window-selection specifications. The paired host minus guest slope contrast was not significant in either the observed or fixed-token analyses. Role-specific embedding–entropy correlations were weak, and turn-level lexical similarity did not increase. Host mean topic divergence tended to exceed guest divergence, but support varied across specifications. Fixed-token document analysis reversed the original entropy ordering between host and guest streams, while external entropy patterns varied across sources. Across the four external sources, the dialogic minus monologic difference in mean dependency distance narrowed from −0.3587 to −0.0661 after sentence-length adjustment. These findings show that declining lexical entropy need not coincide with stronger semantic proximity or local lexical alignment, while external syntactic contrasts require source-aware interpretation.