Adaptive QA Retrieval Using Evidence Spans and Semantic Ranking
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing large language model (LLM)-based question-answering systems face high computational costs, slow response times, and generate incomplete or inaccurate responses, particularly for complex or domain-specific queries, and struggle with retrieving and processing relevant information efficiently.
Innovation Solution
A method involving heuristic-based partitioning and a trained language model for evidence extraction, which partitions text content into paragraphs and extracts relevant evidence spans, performing semantic search to rank candidate passages based on similarity to the input question, using a retriever to enhance relevance and self-containment.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If LLMs are used for question-answering, then the system can understand and generate responses to a broad range of queries, but the computational cost is high and response times are slow
Solution Approach 1:
The patent segments the text content into multiple paragraphs and further divides paragraphs into smaller text spans. This segmentation allows the system to process and search through smaller units of text rather than treating the entire document as one large block, thereby reducing the computational burden and improving response time while maintaining the ability to handle diverse queries.
Solution Approach 2:
The patent extracts relevant evidence spans from the partitioned text content using a trained language model. By extracting only the most relevant portions of text that contain answers to the input question, the system avoids processing the entire text corpus, thus reducing computational cost and response time while maintaining answer accuracy.
2Loss of information
If LLMs retrieve information from external sources, then the system can access broader knowledge, but the retrieval efficiency is low and relevant information may not be found
Solution Approach 1:
The patent partitions text content into paragraphs and evidence spans, creating smaller searchable units. This segmentation enables more efficient retrieval by allowing the system to search through discrete, manageable text segments rather than processing entire documents, thereby improving both retrieval efficiency and the completeness of finding relevant information.
Solution Approach 2:
The patent replaces traditional mechanical search methods with a trained language model that performs semantic matching. Instead of relying on simple keyword matching or manual retrieval processes, the trained model automatically identifies and extracts relevant evidence spans based on semantic understanding, significantly improving retrieval efficiency and accuracy.
3Reliability
If the text content is processed in full, then all relevant information is captured, but the processing complexity and computational resources increase
Solution Approach 1:
The patent divides the text content into paragraphs and further segments paragraphs into smaller text spans. This hierarchical segmentation reduces processing complexity by breaking down large text blocks into manageable units that can be processed independently, while still ensuring that all relevant information is captured through systematic searching of the segmented structure.
Solution Approach 2:
The patent extracts only the relevant evidence spans from the partitioned text content using a trained language model. This extraction process filters out unnecessary information and focuses computational resources on processing only the most relevant text segments, thereby maintaining high reliability in retrieving accurate information while reducing overall processing complexity.
Data Source
AI summary
Methods and systems for retrieving relevant information in response to an input question. The method includes obtaining text content related to the input question and partitioning the content into one or more paragraphs based on predefined rules. The method further involves extracting one or more evidence spans that are relevant to the input question by inputting the text content and the question into a trained language model. A semantic search is then performed on both the paragraphs and the extracted evidence spans, ranking the candidate passages based on their relevance to the input question. Each candidate passage may comprise either a paragraph or an evidence span that addresses the question. The disclosed methods and systems improve the quality and relevance of retrieved information by combining heuristic-based content partitioning with machine learning-based evidence extraction.