Span-Level Evidence Extraction Task

A Span-Level Evidence Extraction Task is an information extraction task that is a span identification task focused on extracting exact text spans serving as supporting evidence for downstream NLU decisions.

• AKA: Token-Level Evidence Extraction, Precise Span Extraction Task.
• Context:
  • It can typically identify Character-Level Boundarys or token-level boundarys.
  • It can typically extract Minimal Sufficient Spans for evidence requirements.
  • It can typically support Multiple Span Selection for complex evidences.
  • It can typically maintain Span Ordering for sequential evidences.
  • It can typically preserve Original Text Form without paraphrasing.
  • ...
  • It can often utilize Sequence Labeling Approaches for span boundary detection.
  • It can often employ Pointer Networks for start-end position prediction.
  • It can often incorporate Span Ranking Algorithms for relevance scoring.
  • It can often require Overlapping Span Handling for nested evidences.
  • ...
  • It can range from being a Single-Span Extraction Task to being a Multi-Span Extraction Task, depending on its span cardinality.
  • It can range from being a Contiguous Span Extraction Task to being a Discontinuous Span Extraction Task, depending on its span continuity requirement.
  • ...
  • It can process Text Documents with evidence annotations.
  • It can output Span Coordinates with confidence scores.
  • It can be solved by a Span-Level Evidence Extraction System.
  • It can be evaluated by a Span Extraction Performance Measure.
  • ...
• Example(s):
  • Question Answering Span Extraction, such as:
    • SQuAD Answer Extraction Task for reading comprehension.
    • Natural Questions Span Task for long-form documents.
  • Named Entity Recognition Tasks viewed as evidence extraction, such as:
    • Clinical Entity Extraction for medical evidences.
    • Legal Entity Extraction for contract terms.
  • Rationale Extraction Tasks, such as:
    • ERASER Rationale Extraction for movie reviews.
    • e-SNLI Evidence Extraction for entailment explanations.
  • ...
• Counter-Example(s):
  • Document Retrieval Tasks, which return entire documents not specific spans.
  • Abstractive Summarization Tasks, which generate new text not extract existing spans.
  • Sentence Classification Tasks, which label whole sentences not sub-sentence spans.
• See: Information Extraction Task, Span Identification Task, Evidence Extraction Task, Sequence Labeling Task.

---