Highlighting Influential Conversation Turns in Intelligent Virtual Assistants
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural network systems, such as Hierarchical Attention Networks (HAN), fail to provide informative real-time visual displays of text and graphics when all conversation turns have uniform attention weights, making it difficult for human representatives to quickly identify the cause of escalation in conversations with intelligent virtual assistants (IVAs).
Innovation Solution
A method that uses sequential long-short-term-memory (LSTM) cells to update weighting values for conversation turns, identifying changes in weight distribution, context, and variation dependencies to create visualizations that highlight influential turns, even when HAN weights are uniform, by calculating a turn weight vector and using perplexity to determine uniformity thresholds.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If Hierarchical Attention Networks are used to weight conversation turns, then the system can identify important turns in the conversation, but when all turns have uniform weights, the system fails to provide informative visual displays making it difficult for human representatives to quickly identify escalation causes
Solution Approach 1:
The patent changes the parameter being measured from static attention weights to dynamic weight changes. Instead of displaying the absolute weight values which may be uniform, the system calculates the change in weights between sequential turns and displays these differences, transforming a uniform parameter distribution into an informative visual representation that highlights influential turns
Solution Approach 2:
The system transitions from a static weight assignment model to a dynamic analysis that tracks how weights evolve across conversation turns. By computing sequential weight changes and using LSTM cells to model temporal dependencies, the system dynamically identifies which turns caused significant shifts in attention, converting uniform static weights into meaningful dynamic information
2Loss of information
If the system displays all conversation turns with their weights, then complete information is provided, but human representatives cannot quickly identify the cause of escalation increasing review time
Solution Approach 1:
The system extracts and highlights only the most influential turns by identifying significant weight changes, rather than displaying all turns uniformly. This extraction of key information maintains completeness for analysis while presenting a condensed visual summary that enables rapid identification of escalation causes
Solution Approach 2:
The patent applies different visual qualities to different conversation turns based on their influence. Turns with significant weight changes are highlighted with distinctive visual properties (such as coloring or emphasis), while less influential turns receive standard treatment, creating a locally differentiated display that guides human attention to critical information
3Loss of information
If the system uses complex sequential analysis with LSTM cells to update weighting values, then more informative visuals are generated, but computation time increases affecting quality of service
Solution Approach 1:
The system performs sequential weight updates only when necessary to capture meaningful changes in conversation dynamics. By using LSTM cells to detect significant transitions rather than computing all possible weight combinations, the system achieves high information quality with reduced computational overhead, processing only the critical sequential dependencies
Data Source
AI summary
Attention weights in a hierarchical attention network indicate the relative importance of portions of a conversation between an individual at one terminal and a computer or a human agent at another terminal. Weighting the portions of the conversation after converting the conversation to a standard text format allows for a computer to graphically highlight, by color, font, or other indicator visible on a graphical user interface, which portions of a conversation led to an escalation of the interaction from an intelligent virtual assistant to a human customer service agent.