<![CDATA[Workplace sexism often appears as indirect, deniable language (e.g., patronizing compliments, competence-doubting questions), making automated detection and organizational response difficult. This study evaluates a transparent, explanation-ready NLP pipeline on the Sexist Workplace Statements (SWS) dataset (1,137 items) with its binary labels: certain sexism vs. ambiguous/neutral. Using the provided fixed stratified split (1,023 train; 114 test), we train a TF–IDF (word 1–2, character 3–5 n-grams) logistic regression baseline and report performance stability across five random seeds. To audit model evidence, sparse token rationales are extracted from linear feature contributions and quantify faithfulness with ERASER-style comprehensiveness (logit drop when rationales are removed) and sufficiency (logit change when only rationales are kept), benchmarked against random-token rationales. The baseline achieves 0.768 ± 0.006 accuracy and 0.759 ± 0.007 macro-F1, with errors concentrated in the ambiguous/neutral class. Faithfulness tests show that model-selected rationales substantially affect the sexism logit (comprehensiveness 1.335 ± 0.001), while remaining insufficient in isolation (|sufficiency| 1.075 ± 0.006). Sanity checks reveal modest sensitivity to gender-term swaps and reduced rationale overlap underweight randomization. Overall, results motivate cautious deployment: explanation-driven auditing can surface shortcut risks and clarify where binary labels blur neutral language and deniable sexism, pointing to future work on finer-grained annotation and human rationale collection.]]>
Copyrights © 2026
