Evidence-Based Text Classification Task

An Evidence-Based Text Classification Task is a NLU task that is a text classification task requiring evidence span extraction to support document-level classification decisions.

• AKA: Evidence-Grounded Classification Task, Span-Supported Text Classification.
• Context:
  • It can typically require Text Span Identification before classification decisions.
  • It can typically enforce Decision Transparency through mandatory evidence requirements.
  • It can typically prevent Unsupported Classification via evidence grounding constraints.
  • It can typically enable Classification Interpretability through highlighted text spans.
  • It can typically support Audit Trail Creation via evidence documentations.
  • ...
  • It can often combine Span Extraction Components with classification components.
  • It can often utilize Joint Evaluation Metrics for label accuracy and evidence quality.
  • It can often integrate Multiple Evidence Types including sentence-level evidence and token-level evidence.
  • It can often require Evidence Completeness for valid classifications.
  • ...
  • It can range from being a Simple Evidence-Based Text Classification Task to being a Complex Evidence-Based Text Classification Task, depending on its evidence requirement complexity.
  • It can range from being a Single-Evidence Text Classification Task to being a Multi-Evidence Text Classification Task, depending on its evidence cardinality.
  • ...
  • It can be evaluated using Joint Classification-Evidence Metrics.
  • It can be solved by an Evidence-Based Text Classification System.
  • It can be measured by an Evidence-Based Classification Performance Measure.
  • It can require Evidence Annotation in training data.
  • ...
• Example(s):
  • Legal Document Classification Tasks requiring clause identification, such as:
    • Contract Risk Classification Task with risk clause highlighting.
    • Legal Compliance Classification Task with regulatory span identification.
  • Scientific Claim Verification Tasks with evidence requirements, such as:
    • FEVER Fact Verification Task requiring Wikipedia evidence sentences.
    • SciFact Verification Task requiring scientific paper evidences.
  • Medical Diagnosis Support Tasks with symptom identification, such as:
    • Clinical Note Classification Task with diagnostic evidence spans.
    • Radiology Report Classification Task with finding localizations.
  • ...
• Counter-Example(s):
  • Black-Box Text Classification Tasks, which lack evidence requirements.
  • End-to-End Neural Classification Tasks, which provide no span identifications.
  • Abstract Text Classification Tasks, which classify without concrete evidences.
• See: Explainable Text Classification, Evidence Extraction Task, Interpretable NLP Task, Span-Based NLU Task.

---