Intent Shift Summarization for Real-Time Multi-Intent Conversations
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing interaction summarization techniques in contact centers suffer from inaccuracies, high latency, and lack of resilience, particularly in handling complex, multi-intent interactions.
Innovation Solution
A system utilizing transformer-based models for intent shift and classification, combined with a generative model, generates real-time summaries by classifying intent shifts and using interaction spans to create accurate prompts, reducing the risk of hallucinations and latency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual summarization by advocates is used, then accuracy can be maintained through human judgment, but productivity is severely reduced due to time-consuming processes
Solution Approach 1:
The patent replaces the mechanical human summarization process with an automated system combining BERT for intent classification, span detection algorithms, and LLMs for summary generation. This substitution maintains accuracy through multiple validation layers while dramatically improving productivity by eliminating manual intervention in the summarization workflow.
Solution Approach 2:
The patent introduces intermediary components including intent classification models and span detection mechanisms that act as mediators between raw interaction data and final summaries. These intermediaries ensure accuracy by validating and structuring information before it reaches the summary generation stage, preventing hallucinations while maintaining high-speed automated processing.
2Productivity
If existing automated summarization systems are used, then productivity is improved through automation, but measurement precision deteriorates due to hallucinations and inaccuracies
Solution Approach 1:
The patent implements feedback mechanisms where span detection results validate LLM output against actual interaction content. The system continuously checks whether generated summaries are supported by detected spans, providing real-time feedback to correct or reject hallucinated content, thereby maintaining high accuracy in automated summarization.
Solution Approach 2:
The patent performs preliminary intent classification and span detection before summary generation. By pre-identifying relevant interaction segments and their intents, the system prepares validated input for LLMs, ensuring that subsequent summary generation is grounded in actual interaction content rather than hallucinated information.
3Adaptability or versatility
If complex multi-intent interactions are processed, then adaptability is improved to handle diverse topics, but device complexity increases due to multiple classification models
Solution Approach 1:
The patent segments complex multi-intent interactions into distinct intent categories using BERT classification, then processes each intent segment separately with span detection. This segmentation allows the system to handle diverse interaction types through specialized processing pipelines for each intent, managing complexity by breaking down multifaceted interactions into manageable components.
Solution Approach 2:
The patent employs a universal architecture where BERT, span detection, and LLM components work together across all interaction types. The same core components handle diverse intents by adjusting classification parameters and span detection thresholds, providing multi-functionality without requiring separate specialized systems for each interaction type.
4Productivity
If real-time summarization is implemented, then productivity is improved through immediate documentation, but measurement precision worsens due to latency in processing
Solution Approach 1:
The patent performs preliminary intent classification and span detection on incoming interaction data before full processing. By pre-processing and identifying key segments in real-time, the system reduces the computational burden during summary generation, minimizing latency while maintaining real-time documentation capability.
Solution Approach 2:
The patent generates partial summaries based on detected spans and intents as interactions unfold, rather than waiting for complete interactions. This partial action approach provides immediate documentation of processed segments, improving real-time productivity while the full summary is finalized with minimal additional latency.
Data Source
AI summary
Techniques for summarizing conversations in real-time are disclosed. The techniques receive streaming data indicating a set of digital interactions between a first user and a second user. The techniques predict a span of time over which a portion of the set of digital interactions are associated with a first intent. The techniques classify a first intent classification associated with the portion of the set of digital interactions. The techniques then generate a first prompt based at least in part on the first intent classification and the portion of the set of digital interactions. The techniques generate a first summary of the set of digital interactions. The summary can then be displayed to a user, e.g., for review and/or editing. The process may repeat within any given digital interaction(s) for multiple intents. These techniques can enhance the accuracy and efficiency of interaction summarization, including summarization of complex, multi-intent interactions.