Generative Language Model for Knowledge Graph Semantic Representation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing cognitive systems face difficulties in effectively representing data stored in knowledge graphs in a natural language question-answering context, lacking an efficient method to utilize semantic relations between entities.
Innovation Solution
A computer-implemented method and system that generates a knowledge graph from a corpus, training a generative language model to associate related entities and produce natural language passages representing the relationships between them, using source passages to refine the representation of edges in the graph.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If knowledge graph data is stored in structured format for efficient processing, then data utilization efficiency is improved, but natural language representation capability deteriorates
Solution Approach 1:
A generative language model is introduced as an intermediary component between the structured knowledge graph and natural language output. The model takes structured entity-relationship data from the knowledge graph and generates natural language passages that represent these relationships, thereby enabling efficient structured processing while maintaining natural language representation capability.
2Manufacturing precision
If semantic relations are extracted from unstructured corpus to build knowledge graph, then data structure quality is improved, but information loss increases
Solution Approach 1:
The system incorporates feedback mechanisms where the generative language model is trained on source passages from the original unstructured corpus. This training process allows the model to learn and preserve nuanced semantic information from the source text while organizing it into structured knowledge graph format, thereby reducing information loss during the structuring process.
3Measurement precision
If generative language model is trained on knowledge graph edges, then natural language generation accuracy is improved, but computational complexity increases
Solution Approach 1:
The training process is segmented into distinct phases: first, knowledge graph extraction from the corpus; second, collecting source passages associated with each edge; third, training the generative language model on these organized passages. This segmentation allows the system to manage computational complexity by processing data in manageable stages rather than attempting to train on the entire corpus at once.
Data Source
AI summary
A computer-implemented method includes generating from a corpus a knowledge graph comprising a plurality of nodes interconnected by a plurality of edges. Each node of the plurality of nodes represents an entity extracted from the corpus. Each edge of the plurality of edges represents a relationship between corresponding entities extracted from the corpus. The knowledge graph includes a plurality of source passages extracted from the corpus from which the plurality of edges was generated. A generative language model is trained, for each edge of the plurality of edges, to associate two or more related entities by utilizing the knowledge graph. Using the trained generative language model, one or more passages representing the edge of the knowledge graph are generated for each edge of the knowledge graph.