Rationale-Guided Text Classification Task

A Rationale-Guided Text Classification Task is a text classification task that is an interpretable NLP task requiring rationale selection before classification decisions to ensure decision transparency.

• AKA: Select-Then-Classify Task, Rationale-Based Classification Task.
• Context:
  • It can typically enforce Two-Stage Processing with selection stage and classification stage.
  • It can typically require Minimal Rationale Extraction for efficiency constraints.
  • It can typically ensure Classification Faithfulness through architectural constraints.
  • It can typically support Human Interpretability via discrete rationale selection.
  • It can typically enable Decision Debugging through rationale inspection.
  • ...
  • It can often limit Rationale Length to percentage of input.
  • It can often require Contiguous Rationales or allow discontinuous rationales.
  • It can often incorporate Rationale Plausibility in evaluation metrics.
  • It can often balance Rationale Sufficiency with rationale comprehensiveness.
  • ...
  • It can range from being a Hard Rationale Selection Task to being a Soft Rationale Selection Task, depending on its selection mechanism.
  • It can range from being a Single-Rationale Classification Task to being a Multi-Rationale Classification Task, depending on its rationale cardinality.
  • ...
  • It can be solved by a Rationale-Guided Text Classification System.
  • It can be evaluated by a Rationale-Based Classification Measure.
  • It can require Rationale Annotations in training data.
  • It can support Model Interpretability Research.
  • ...
• Example(s):
  • Movie Review Classification with sentiment rationales, such as:
    • IMDB Rationale Classification extracting opinion phrases.
    • Rotten Tomatoes Classification with critical excerpts.
  • Medical Diagnosis Classification with symptom rationales, such as:
    • Clinical Note Classification selecting diagnostic indicators.
    • Radiology Report Classification highlighting finding descriptions.
  • Legal Document Classification with clause rationales, such as:
    • Contract Risk Classification extracting risk indicators.
    • Patent Classification selecting claim elements.
  • ...
• Counter-Example(s):
  • End-to-End Classification Tasks, which lack rationale requirements.
  • Black-Box Classification Tasks, which provide no intermediate selection.
  • Full-Document Classification Tasks, which use entire input not selected rationales.
• See: Interpretable Text Classification, Evidence-Based Classification Task, Explainable NLP Task, Select-Then-Predict Task.

---