Interpretable Discriminative Text Representations via Agreement and Label Disentanglement

In three linesLFD (LLM-assisted Feature Discovery) method generates interpretable text representations via inter-annotator agreement (Cohen's κ) and label disentanglement. Validated on 10 text classification tasks: features clearer and less label-entangled than bottleneck baseline, confirmed by human audit (232 raters).

Summary generated by Claude — human-verified