Selective Text Classification Task

A Selective Text Classification Task is a text classification task that is an interpretable NLP task requiring identification and use of only the most relevant text spans for making classification decisions.

• AKA: Sparse Classification Task, Selective Attention Classification.
• Context:
  • It can typically enforce Sparsity Constraints on input selection.
  • It can typically require Relevance Filtering before classification.
  • It can typically promote Model Efficiency through reduced input.
  • It can typically enable Decision Transparency via selection visibility.
  • It can typically support Noise Robustness by ignoring irrelevant parts.
  • ...
  • It can often limit Selection Percentage to small fraction.
  • It can often employ Hard Attention Mechanisms for discrete selection.
  • It can often incorporate Information Bottlenecks in architecture.
  • It can often balance Selection Quality with classification accuracy.
  • ...
  • It can range from being a Fixed-Length Selection Task to being a Variable-Length Selection Task, depending on its selection constraint.
  • It can range from being a Token-Level Selection Task to being a Sentence-Level Selection Task, depending on its selection granularity.
  • ...
  • It can process Noisy Text Documents requiring focused classification.
  • It can output Classification Labels with selected spans.
  • It can be solved by a Selective Text Classification System.
  • It can be evaluated by a Selective Classification Performance Measure.
  • ...
• Example(s):
  • Document Classification with key sentence selection, such as:
    • News Categorization selecting headline and lead paragraph.
    • Email Classification focusing on subject and key phrases.
  • Review Classification with opinion extraction, such as:
    • Product Review Classification selecting sentiment phrases.
    • Movie Review Classification extracting critical excerpts.
  • Medical Text Classification with symptom selection, such as:
    • Diagnosis Classification focusing on clinical findings.
    • Triage Classification selecting urgent indicators.
  • ...
• Counter-Example(s):
  • Full-Document Classification Tasks, which use entire input.
  • Dense Classification Tasks, which process all tokens equally.
  • Black-Box Classification Tasks, which hide selection process.
• See: Sparse Text Processing, Interpretable Classification, Attention-Based Classification, Efficient NLP Task.

---