Multi-Hop Evidence Retrieval Task

A Multi-Hop Evidence Retrieval Task is an evidence retrieval task that is a complex NLU task requiring sequential evidence extraction across multiple documents to support fact verification or question answering.

• AKA: Multi-Step Evidence Task, Chain Evidence Retrieval Task.
• Context:
  • It can typically require Reasoning Chain Construction across document boundarys.
  • It can typically involve Evidence Dependency Resolution between retrieval steps.
  • It can typically necessitate Cross-Document Coreference for entity linking.
  • It can typically demand Inference Path Tracking through evidence sequences.
  • It can typically support Complex Query Resolution via iterative retrieval.
  • ...
  • It can often incorporate Bridge Entity Recognition connecting evidence pieces.
  • It can often utilize Hierarchical Retrieval Strategyes for efficiency.
  • It can often require Evidence Ordering for logical flow.
  • It can often employ Confidence Propagation across retrieval hops.
  • ...
  • It can range from being a Two-Hop Retrieval Task to being a Many-Hop Retrieval Task, depending on its reasoning depth.
  • It can range from being a Homogeneous Corpus Task to being a Heterogeneous Corpus Task, depending on its source diversity.
  • ...
  • It can process Multi-Document Collections with linked information.
  • It can output Evidence Chains with supporting documents.
  • It can be solved by a Multi-Hop Evidence Retrieval System.
  • It can be evaluated by a Multi-Hop Retrieval Performance Measure.
  • ...
• Example(s):
  • HotpotQA Task requiring two-hop reasoning over Wikipedia articles.
  • Multi-hop FEVER Task needing multiple evidence sentences for claim verification.
  • WikiHop Task extracting answers through entity chains.
  • Complex Question Answering over knowledge graphs with text evidence.
  • Legal Precedent Retrieval tracing case citations across documents.
  • ...
• Counter-Example(s):
  • Single-Document QA Tasks, which find answers in one document.
  • Direct Retrieval Tasks, which return relevant documents without chaining.
  • Extractive Summarization Tasks, which process documents independently.
• See: Evidence Retrieval Task, Complex Question Answering, Multi-Document Processing, Reasoning Task.

---