Multi-Hop Evidence Retrieval System
Jump to navigation
Jump to search
A Multi-Hop Evidence Retrieval System is an evidence retrieval system that is a complex reasoning system implementing iterative retrieval algorithms to solve multi-hop evidence retrieval tasks.
- AKA: Chain Reasoning System, Multi-Step Evidence Extractor.
- Context:
- It can typically implement Iterative Retrieval Modules for sequential search.
- It can typically maintain Reasoning State Trackers across retrieval hops.
- It can typically employ Entity Linkers for cross-document connections.
- It can typically utilize Evidence Aggregators for chain construction.
- It can typically provide Reasoning Path Visualizations for interpretability.
- ...
- It can often incorporate Graph Neural Networks for reasoning representation.
- It can often use Reinforcement Learning for retrieval policy.
- It can often implement Beam Search for multiple reasoning paths.
- It can often employ Attention Mechanisms for evidence selection.
- ...
- It can range from being a Pipeline Multi-Hop System to being an End-to-End Multi-Hop System, depending on its architecture design.
- It can range from being a Sparse Retrieval System to being a Dense Retrieval System, depending on its retrieval method.
- ...
- It can process Document Collections with complex querys.
- It can output Evidence Chains with confidence scores.
- It can integrate with Question Answering Systems for answer generation.
- It can be evaluated by Multi-Hop Retrieval Performance Measures.
- ...
- Example(s):
- HotpotQA Baseline Systems using TF-IDF and graph traversal.
- Dense Passage Retriever adapted for iterative retrieval.
- REALM System with multi-hop capability for knowledge-intensive tasks.
- Graph-Based Reasoning Systems over document networks.
- Neural Multi-Hop Retrievers with learned routing.
- ...
- Counter-Example(s):
- Single-Pass Retrieval Systems, which perform one-shot retrieval.
- Document Ranking Systems, which order documents without chaining.
- Keyword Search Systems, which lack reasoning capability.
- See: Complex Reasoning System, Evidence Retrieval System, Multi-Document System, Iterative Search System.