{"id":"W4389518176","doi":"10.18653/v1/2023.blackboxnlp-1.27","title":"Systematic Generalization by Finetuning? Analyzing Pretrained Language Models Using Constituency Tests","year":2023,"lang":"en","type":"article","venue":"","topic":"Topic Modeling","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"ca_institutions":"McGill University","funders":"Samsung; Compute Canada","keywords":"Sentence; Computer science; Natural language processing; Artificial intelligence; Generalization; Copying; Replicate; Language model; Word (group theory); Linguistics","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":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.004760685,0.001230933,0.0008682279,0.000375541,0.0002412678,0.001197479,0.001408522,0.001021905,0.001622195],"category_scores_gemma":[0.03241166,0.0007066416,0.0007733263,0.0002430481,0.0009455758,0.003540392,0.001438998,0.003208415,0.0004106197],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0007450063,"about_ca_system_score_gemma":0.0006100066,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.002410436,"about_ca_topic_score_gemma":0.002742335,"domain_scores_codex":[0.9981381,0.0008350905,0.0001637234,0.0005369915,0.0001862332,0.0001398601],"domain_scores_gemma":[0.9790243,0.01303879,0.001118678,0.005654789,0.0007560353,0.0004074771],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.001371251,0.0009706913,0.05135992,0.0005328093,0.001060845,0.0004380759,0.001484675,0.6167787,0.1137583,0.004074625,0.003244278,0.2049258],"study_design_scores_gemma":[0.0000627179,0.0005666709,0.0103354,0.00003009685,0.0001266389,0.00007568657,0.0001923037,0.9466438,0.02929622,0.01170314,0.0009073805,0.00005994905],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9256384,0.0002648782,0.06987706,0.0004114703,0.00006677341,0.0001339591,0.0002801518,0.00152317,0.001804212],"genre_scores_gemma":[0.9869764,0.0000553822,0.01177008,0.0001415139,0.00001239848,0.0001095771,0.0003944689,0.000177556,0.0003626519],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.004760685,"threshold_uncertainty_score":0.02517718,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.03343015652090418,"score_gpt":0.2776285085134947,"score_spread":0.2441983519925905,"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."}}