Many-Hop Fact Extraction for Robust Claim Verification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing fact extraction and verification systems struggle with multi-hop reasoning, often relying on single-source information and are susceptible to adversarial evaluations, limiting their effectiveness in real-world scenarios with complex claims.
Innovation Solution
A machine learning system and dataset (HoVer) that extracts facts from multiple documents and verifies claims using a 4-stage architecture, including document and sentence retrieval, and natural language inference, to handle complex multi-hop reasoning scenarios.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If single-source information retrieval is used, then system complexity is reduced, but verification accuracy deteriorates due to inability to handle multi-hop reasoning
Solution Approach 1:
The system segments the verification process into multiple independent stages: document retrieval, sentence selection, fact extraction, and claim verification. Each stage processes information from multiple sources independently, then integrates results to achieve accurate multi-hop reasoning without overwhelming system complexity
Solution Approach 2:
The system transitions from single-source to multi-source information retrieval by adding the dimension of multiple documents and sentences. This dimensional expansion enables the system to capture multi-hop reasoning paths while maintaining manageable complexity through structured processing
2Speed
If word-matching reasoning shortcuts are used, then processing speed is improved, but reliability deteriorates due to susceptibility to adversarial evaluations
Solution Approach 1:
The system introduces machine learning models as intermediaries between word-matching and final verification. These models act as mediators that process retrieved sentences and facts through semantic understanding, preventing adversarial evaluations from exploiting simple word-matching shortcuts while maintaining reasonable processing speed
Solution Approach 2:
The system replaces mechanical word-matching operations with machine learning-based semantic analysis. This substitution eliminates reliance on brittle keyword matching that is vulnerable to adversarial attacks, while the automated ML-based processing maintains efficient speed
3Measurement precision
If multi-document retrieval is implemented, then fact extraction accuracy is improved, but device complexity increases due to handling multiple sources
Solution Approach 1:
The system segments multi-document processing into distinct stages: retrieving relevant documents, selecting pertinent sentences from each document, extracting facts, and verifying claims. This segmentation enables accurate handling of multiple sources while keeping each stage's complexity manageable and modular
Solution Approach 2:
The system employs universal machine learning models that can process information from any document source through the same pipeline. This multi-functionality allows the system to handle diverse multi-document inputs with a single unified architecture, avoiding the need for source-specific processing complexity
4Ease of operation
If single-hop verification is used, then ease of operation is maintained, but adaptability deteriorates for complex real-world claims
Solution Approach 1:
The system implements dynamic verification depth, automatically adapting the number of reasoning hops based on claim complexity. For simple claims, it performs single-hop verification for ease of operation; for complex real-world claims, it dynamically extends to multi-hop reasoning, maintaining both operational simplicity and adaptability
Data Source
AI summary
Machine learning (ML) systems and methods for fact extraction and claim verification are provided. The system receives a claim and retrieves a document from a dataset. The document has a first relatedness score higher than a first threshold, which indicates that ML models of the system determine that the document is most likely to be relevant to the claim. The dataset includes supporting documents and claims including a first group of claims supported by facts from more than two supporting documents and a second group of claims not supported by the supporting documents. The system selects a set of sentences from the document. The set of sentences have second relatedness scores higher than a second threshold, which indicate that the ML models determine that the set of sentences are most likely to be relevant to the claim. The system determines whether the claim includes facts from the set of sentences.