Knowledge Graph Summaries for Efficient Natural Language Generation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing natural language generation systems, particularly in question-answering (QA) systems, face challenges with high computational resource requirements, limited domain-specific information, and inefficient retrieval of relevant information due to sparse or dense data, leading to suboptimal answer quality and resource consumption.
Innovation Solution
A knowledge graph incorporating textual summaries is used to provide improved input context to language models, utilizing a hierarchical structure with document summaries and textual passages, enabling targeted retrieval of information with reduced computational and memory demands.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional natural language generation systems process dense or sparse data directly, then they can provide domain-specific information, but they consume high computational resources and memory
Solution Approach 1:
The patent segments the knowledge graph into hierarchical levels (document level, section level, passage level) and retrieves only relevant segments based on the query. This segmentation allows the system to process domain-specific information without loading entire documents or datasets into memory, thereby reducing computational resource consumption while maintaining answer quality.
Solution Approach 2:
The patent performs preliminary action by pre-computing and storing embeddings for passages, sections, and documents in the knowledge graph during an offline phase. This preliminary processing enables the online query phase to efficiently retrieve relevant information without performing heavy computations in real-time, thus reducing online computational resource consumption while preserving answer quality.
2Loss of information
If the system retrieves comprehensive textual passages from the knowledge graph, then information retrieval completeness improves, but retrieval efficiency and processing speed decrease
Solution Approach 1:
The patent applies local quality by differentiating the retrieval strategy at different hierarchical levels. At the passage level, it retrieves highly relevant local passages; at the section level, it retrieves broader contextual sections; and at the document level, it retrieves relevant documents. This localized approach ensures comprehensive information retrieval while optimizing retrieval efficiency by not retrieving unnecessary data at each level.
Solution Approach 2:
The patent introduces a hierarchical dimension to the retrieval process, organizing knowledge at multiple levels (passage, section, document). This dimensional change allows the system to retrieve information comprehensively across levels while improving efficiency by filtering and prioritizing retrieval at each hierarchical stage, avoiding the need to retrieve all data simultaneously.
3Reliability
If the language model processes multiple textual passages and summaries jointly, then contextual relevance of answers improves, but computational complexity increases
Solution Approach 1:
The patent merges multiple textual passages and summaries at different hierarchical levels into a unified context for the language model. By combining relevant passages, sections, and documents that are retrieved from the knowledge graph, the system provides comprehensive contextual information to the model, improving answer contextual relevance while managing computational complexity through selective merging of only necessary components.
Solution Approach 2:
The patent introduces an intermediary layer (the knowledge graph with hierarchical structure and embeddings) between the query and the language model. This intermediary pre-processes and organizes domain-specific information, filtering and structuring data before presenting it to the model. This mediation reduces the computational complexity the model would otherwise face by providing pre-organized, relevant context rather than raw unprocessed data.
Data Source
AI summary
Some embodiments relate to receiving a natural-language textual sequence representing; retrieving, from a knowledge graph, a first textual passage and a second textual passage based on rankings with respect to the natural-language textual sequence, a first textual summary summarizing textual information in a first vicinity of the first textual passage, and a second textual summary summarizing textual information in a vicinity of the second textual passage; obtaining the textual output in response to the textual input using a language model by encoding a first intermediate output based on the natural-language textual sequence, the first textual passage, and the first textual summary, encoding a second intermediate output based on the natural language textual sequence, the second textual passage, and the second textual summary, and decoding a concatenation of the first intermediate output and the second intermediate output; and providing an output to a user based on the textual output.