Hybrid Extractive-Abstractive Summarization for Multi-Party Transcripts

The patent combines extractive summarization (which selects verbatim sentences from the source) with abstractive summarization (which generates new sentences) in a hybrid pipeline. The extractive component ensures factual accuracy by selecting only sentences that exist in the source text, while the abstractive component improves coherence and context capture. This merging resolves the contradiction by leveraging the strengths of both approaches: extractive methods prevent hallucinations while abstractive methods enhance context understanding.

The patent introduces an intermediary verification step where a separate model evaluates the output of the abstractive summarization component for hallucinated content. This intermediary layer acts as a filter that identifies and removes fabricated information while preserving the contextual benefits of abstractive methods. The verification model serves as a mediator between the creative abstractive process and the final output, ensuring reliability.

The hybrid summarization system merges extractive and abstractive methods, where the extractive component selects accurate verbatim sentences and the abstractive component generates coherent connecting text. This combination allows the system to maintain the accuracy benefits of extractive methods while adding the contextual coherence benefits of abstractive methods, resolving the contradiction between accuracy and contextual information.

The patent segments the summarization task into distinct components: an extractive summarization module that handles accuracy-critical sentence selection, and an abstractive summarization module that handles coherence and context. By dividing the task into segments with specialized functions, each component can optimize for its strength while the integrated system achieves both accuracy and contextual richness.

The complex task of summarizing multi-party interactions is segmented into multiple specialized processing stages: utterance segmentation, speaker identification, extractive summarization, abstractive summarization, and verification. Each segment handles a specific aspect of the problem, making the overall complex task manageable and automatable. This segmentation enables high productivity by allowing parallel processing of different aspects while maintaining systematic control over the complexity.

The patent employs universal machine learning models that can handle multiple functions within the summarization pipeline. For example, the same transformer-based architecture is used for both extractive and abstractive components, and the verification model serves multiple purposes including hallucination detection and coherence checking. This multi-functionality reduces the overall system complexity by reusing proven models across different tasks rather than requiring specialized models for each function.

The patent extracts only the most relevant and non-redundant information from multi-party interactions using the extractive summarization component. By selecting key sentences that capture essential information while automatically filtering out redundant content, the system achieves information completeness without proportionally increasing processing time. The extraction process inherently eliminates redundancy by selecting only the most important sentences.

The system applies partial abstractive summarization rather than complete transformation of all content. The abstractive component processes only the extractive summary output and focuses on adding contextual connections where needed, rather than completely regenerating the entire summary. This partial action approach maintains information completeness while significantly reducing processing time compared to full abstractive summarization.