Evidence Diffusion for Candidate Answer Ranking in QA Systems
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current Question Answering (QA) systems fail to effectively utilize relationships among candidate answers, leading to incorrect rankings and reduced precision in answering natural language questions, as they typically consider candidates independently and do not diffuse evidence between them.
Innovation Solution
The proposed method identifies relationships between candidate answers and transfers evidence between them, using structured knowledge bases and shallow lexical knowledge to re-rank candidates based on enhanced feature scores, thereby improving the confidence scores of correct answers.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If QA systems consider candidate answers independently without exploring relationships among them, then the system complexity is reduced and processing is simpler, but the precision and accuracy of answer ranking deteriorates
Solution Approach 1:
The patent merges the evaluation of candidate answers by introducing evidence diffusion that combines evidence from multiple candidates. Instead of evaluating each candidate independently, the system aggregates evidence across candidates through diffusion processes, allowing the correct answer to accumulate sufficient evidence from related candidates to achieve accurate ranking.
Solution Approach 2:
The patent introduces evidence as an intermediary element that mediates between candidate answers. Evidence diffuses from one candidate to another through identified relationships, acting as a carrier that transfers supporting information across the candidate set. This intermediary mechanism enables indirect evaluation where candidates can benefit from evidence originating in related candidates.
2Reliability
If QA systems transfer evidence between candidate answers based on identified relationships, then the accuracy of answer identification is improved, but the computational time and processing complexity increase
Solution Approach 1:
The patent applies partial action by performing evidence diffusion selectively rather than exhaustively. The system identifies relationships among candidates and diffuses evidence along these identified paths, focusing computational effort on the most promising connections. This partial approach achieves sufficient accuracy improvement without the prohibitive cost of examining all possible candidate relationships.
Solution Approach 2:
The patent performs preliminary relationship identification among candidate answers before conducting evidence diffusion. By pre-establishing the relationship structure among candidates, the system prepares the framework for efficient evidence transfer. This preliminary action organizes the computational workload in advance, enabling faster execution of the actual evidence diffusion process.
3Measurement precision
If QA systems use structured knowledge bases and shallow lexical knowledge to identify relationships among candidates, then the precision of evidence transfer is improved, but the device complexity and data processing requirements increase
Solution Approach 1:
The patent applies local quality by using different types of knowledge resources for different aspects of relationship identification. Structured knowledge bases provide formal, structured relationships while shallow lexical knowledge handles informal linguistic connections. This localized application of different knowledge qualities enables precise evidence transfer without requiring a single complex system to handle all relationship types uniformly.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Diffusing evidence among candidate answers during question answering may identify a relationship between a first candidate answer and a second candidate answer, wherein the candidate answers are generated by a question-answering computer process, the candidate answers have associated supporting evidence, and the candidate answers have associated confidence scores. All or some of the evidence may be transferred from the first candidate answer to the second candidate answer based on the identified relationship. A new confidence score may be computed for the second candidate answer based on the transferred evidence.