{"id":"W3206995700","doi":"10.18653/v1/2022.findings-emnlp.125","title":"Evaluating the Faithfulness of Importance Measures in NLP by Recursively Masking Allegedly Important Tokens and Retraining","year":2022,"lang":"en","type":"preprint","venue":"","topic":"Explainable Artificial Intelligence (XAI)","field":"Computer Science","cited_by":16,"is_retracted":false,"has_abstract":true,"ca_institutions":"McGill University; Polytechnique Montréal; Mila - Quebec Artificial Intelligence Institute","funders":"Compute Canada; Natural Sciences and Engineering Research Council of Canada; Canadian Institute for Advanced Research","keywords":"Masking (illustration); Computer science; Retraining; Metric (unit); Artificial intelligence; Natural language processing; Task (project management); Machine learning; Closed captioning; Language model","routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false,"invisible_to_affiliation_only":false},"retraction":null,"screen":null,"direct_labels":[],"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.006946386,0.0003910784,0.0005789277,0.0002174148,0.0003246935,0.0002628124,0.002385164,0.0001868604,0.0001411328],"category_scores_gemma":[0.0009384563,0.0003249932,0.000125404,0.0006686735,0.000179386,0.0003654989,0.002920431,0.001303857,0.000001292203],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0002077833,"about_ca_system_score_gemma":0.0004775528,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.001362834,"about_ca_topic_score_gemma":0.000882242,"domain_scores_codex":[0.9950538,0.0005694288,0.001382343,0.001182755,0.00119596,0.0006156578],"domain_scores_gemma":[0.9967172,0.0006348664,0.0009951453,0.001312717,0.0002452396,0.00009486135],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.0003008702,0.0007120951,0.08711891,0.00104845,0.0005770883,0.0006767751,0.1619515,0.06079702,0.02699578,0.3532073,0.007151874,0.2994623],"study_design_scores_gemma":[0.0004213253,0.0006881657,0.00576121,0.0007389407,0.00006503685,0.00007229154,0.02302357,0.8086732,0.03409855,0.1234588,0.001190061,0.001808853],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.7878067,0.007119996,0.1859656,0.004752359,0.00154909,0.00200898,0.00005118652,0.0002704599,0.01047566],"genre_scores_gemma":[0.982015,0.000183681,0.01700702,0.0003318099,0.00004613064,0.0001930856,0.00001017354,0.00003157822,0.0001815515],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.7478762,"threshold_uncertainty_score":0.9999202,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.1136404728347689,"score_gpt":0.3655468697069478,"score_spread":0.2519063968721789,"validation_status":"score_only:v0-immature-baseline","note":"Baseline scores from an immature model (maturity gate not passed). Scores rank; they never assert a category."}}