Long-Range Dependency Modeling and Decision Point Summarization for Large Language Models in Dialogue and Meeting Scenarios

Abstract

In highly interactive text scenarios such as dialogues and meetings, maintaining long-term stability and accurately condensing key decision points are the core challenges in generating decision-oriented summaries. Meeting transcriptions typically feature long rounds, frequent topic shifts, omissions of references, and scattered key information, easily leading to omissions of key points, missing conditions, or drifting conclusions. To address these issues, a representation framework based on rounds is constructed. Contextual continuity is maintained through dialogue structuring and long text segmentation, and cross-round information is accumulated using global contextual memory to improve consistent understanding of key entities, constraint updates, and discussion flow. Building upon this, an explicit scoring and gating mechanism oriented towards decision points is introduced to focus on key rounds that carry out conclusion confirmation, solution selection, and action determination. Combined with attention selection, weighted aggregation of evidence fragments is achieved, making the summary output more focused on the complete organization of decision elements such as conclusions, conditions, and action items. To enhance interpretability and auditability, a retrospective evaluation approach is further adopted. This approach establishes a correspondence between the summary content and original evidence fragments, characterizing the concentration and dispersion of evidence alignment. The contribution distribution between summary sentences and dialogue rounds is then visualized. Comparative evaluation shows that this approach is competitive in overall metrics, helping to more stably aggregate and compress decision-related information in long meeting contexts, and improving the usability of decision-oriented summaries in organizational collaboration and knowledge accumulation.