Conversation Graph Generation for Virtual Agent Knowledge Retrieval
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current natural language understanding systems for virtual agents are unable to effectively utilize the knowledge from previous conversations for similar tasks, leading to inefficiencies in processing and responding to user queries.
Innovation Solution
A system and method for generating a conversation graph that represents a collection of conversations about a specific topic by assigning word sequences to nodes and connecting them with transitional paths, leveraging spoken language understanding and dialogue act category classification to create a comprehensive knowledge base for improved real-time conversations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If a dense and complex collection of conversation data is stored in the knowledge base, then the comprehensiveness of knowledge is improved, but the complexity of accessing and evaluating the knowledge base increases
Solution Approach 1:
The patent segments the dense conversation data into structured conversation graphs with discrete nodes representing utterances and edges representing transitions. This segmentation organizes the knowledge base into manageable units that can be efficiently accessed and evaluated, resolving the contradiction between comprehensiveness and accessibility complexity.
Solution Approach 2:
The patent transforms the flat, dense collection of conversation data into a multi-dimensional graph structure with nodes, edges, and hierarchical relationships. This dimensional transformation enables efficient navigation and access to specific knowledge portions while maintaining the comprehensive nature of the original data collection.
2Measurement precision
If the virtual agent evaluates a large database of knowledge about a specific task, then the accuracy of responses is improved, but the computational resources and computation time increase
Solution Approach 1:
The patent performs preliminary action by pre-processing conversations into structured graph representations during offline training. This preliminary structuring enables the virtual agent to quickly query and retrieve relevant knowledge during online interactions without performing complex evaluations in real-time, thus maintaining accuracy while reducing computation time.
Solution Approach 2:
The patent extracts only the necessary portions of the knowledge base that are relevant to the current user query by navigating the conversation graph structure. This selective extraction avoids evaluating the entire large database, reducing computational resources and time while maintaining response accuracy through targeted knowledge retrieval.
3Device complexity
If the virtual agent uses traditional natural language understanding systems, then the simplicity of the system is maintained, but the ability to utilize knowledge from previous conversations is reduced
Solution Approach 1:
The patent introduces conversation graphs as an intermediary structure between the user query and the knowledge base. This intermediary representation enables the virtual agent to effectively utilize knowledge from previous conversations by providing a structured pathway for knowledge retrieval, while maintaining relative system simplicity through the use of graph-based representations rather than complex neural architectures.
Data Source
AI summary
A system and method for generating a conversation graph for a group of related conversations is disclosed. The system and method use an iterative process of clustering multiple conversations into labeled clusters having similar user intents. The labeled clusters may be used to train a virtual agent to classify the conversational intent of a conversation. Utterances by the agent and/or customer in each conversation from a group of conversations about a similar task or goal can be processed and the dialogue categorized. The resultant classifications are used to represent the many conversations in a single graph by a plurality of nodes interconnected by transitional paths that indicate the conversation flow.