Knowledge Graph Summaries for Low-Resource QA Answer Generation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing natural language generation systems, particularly Question-Answer (QA) systems, face challenges due to high computational resource requirements, limited input size, and inefficiencies in retrieving relevant information, making it difficult to provide high-quality answers without substantial resources and time for training and fine-tuning.
Innovation Solution
A knowledge graph incorporating textual summaries is generated to provide improved input context to language models, utilizing a decoder-only model to produce textual passages and relationships, and summaries organized in a tree format, enabling efficient retrieval and generation of answers.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If existing QA systems use traditional language models with full context input, then answer quality is maintained, but computational resource requirements and memory usage increase significantly
Solution Approach 1:
The patent segments the input context into two distinct parts: textual passages (raw text) and textual summaries (condensed representations). This segmentation allows the system to process and store both forms separately, enabling efficient retrieval where summaries provide quick context while passages provide detailed information when needed, thus reducing overall computational load while maintaining answer quality.
Solution Approach 2:
The patent introduces textual summaries as an intermediary layer between the original corpus and the language model input. These summaries act as mediators that condense large amounts of information into compact forms, allowing the system to provide context-rich input to the language model without directly processing the full original text, thereby reducing computational resources while preserving answer quality.
2Loss of information
If traditional QA systems process large corpora directly, then comprehensive information is available, but retrieval efficiency and response time decrease
Solution Approach 1:
The patent applies preliminary action by generating textual summaries of the corpus before the actual QA process. These summaries are pre-computed and stored alongside the original passages, creating a prepared index that enables rapid retrieval. When a question is asked, the system can quickly search through summaries to identify relevant passages without having to process the entire corpus in real-time, thus maintaining comprehensive information availability while dramatically improving retrieval efficiency.
3Measurement precision
If language models are trained and fine-tuned extensively, then answer accuracy improves, but training time and computational cost increase
Solution Approach 1:
The patent implements self-service by enabling the language model to operate effectively with a dual-input structure (passages and summaries) that provides rich context without requiring extensive domain-specific fine-tuning. The system's architecture itself—combining condensed summaries with original passages—serves as a form of self-adaptation that allows the model to achieve high answer accuracy using its general capabilities, thereby avoiding the need for time-consuming and resource-intensive fine-tuning processes.
4Loss of information
If full textual context is provided to the language model, then context richness is maximized, but input size limitations are exceeded and processing efficiency decreases
Solution Approach 1:
The patent adds another dimension to the input structure by introducing a hierarchical organization with two levels: textual summaries (higher-level condensed information) and textual passages (lower-level detailed information). This dimensional change allows the system to provide comprehensive context within constrained input sizes by selectively combining summary and passage content, effectively expanding the information density without proportionally increasing the raw input size.
Data Source
AI summary
Some techniques relate to generating a knowledge graph including textual passages and textual summaries usable for producing an answer to a question relating to a domain. A corpus of textual information is received. Textual passages and descriptions of associations between textual passages from a first language model are obtained by providing at least a portion of the corpus of textual information to the first language model. Textual summaries corresponding to textual passages are obtained from a second language model. A knowledge graph is generated based on the textual passages, descriptions of associations between the textual passages, and textual summaries. The knowledge graph is stored in a non-transitory computer readable medium.