Explainable Span Extraction Task

An Explainable Span Extraction Task is a span extraction task that is an interpretable NLP task requiring explicit mappings between extracted spans and the NLU decisions they support.

• AKA: Interpretable Span Extraction Task, Span-Decision Mapping Task.
• Context:
  • It can typically require Span-Purpose Annotations linking spans to decision.
  • It can typically demand Extraction Justifications for each span.
  • It can typically enforce Traceable Connections between evidence and outcome.
  • It can typically support Decision Debugging through span inspection.
  • It can typically enable Error Analysis via span-decision alignment.
  • ...
  • It can often involve Multi-Level Explanations from token to document level.
  • It can often require Confidence Scores for span-decision links.
  • It can often incorporate Alternative Spans with ranking explanations.
  • It can often support Interactive Exploration of span relevance.
  • ...
  • It can range from being a Single-Decision Span Task to being a Multi-Decision Span Task, depending on its decision complexity.
  • It can range from being a Direct Mapping Task to being a Inferential Mapping Task, depending on its explanation depth.
  • ...
  • It can process Text Documents requiring interpretable extraction.
  • It can output Span-Decision Pairs with explanation links.
  • It can be solved by an Explainable Span Extraction System.
  • It can be evaluated by an Explainable Extraction Performance Measure.
  • ...
• Example(s):
  • Legal Evidence Extraction with claim support mapping, such as:
    • Contract Term Extraction linking to risk assessment.
    • Statute Citation Extraction supporting legal argument.
  • Medical Span Extraction with diagnosis mapping, such as:
    • Symptom Extraction linked to disease prediction.
    • Lab Value Extraction supporting treatment decision.
  • Financial Document Extraction with decision support, such as:
    • Risk Factor Extraction for investment recommendation.
    • Performance Metric Extraction for rating decision.
  • ...
• Counter-Example(s):
  • Blind Span Extraction Tasks, which extract without decision context.
  • Post-hoc Explanation Tasks, which explain after extraction complete.
  • Generic NER Tasks, which identify entitys without purpose mapping.
• See: Interpretable Information Extraction, Evidence-Based NLP Task, Explainable AI Task, Span-Purpose Alignment.

---