Robust Semantic Role Labeling for NominalsIn briefArchitectureOur contributionSlide 5Data analysisConclusionsRobust Semantic Role Labeling for NominalsRobert MunroAman NaimatIn brief•Created a system for Nominal Semantic Role Labeling•Useful for Information Extraction and Q&A:•An ExampleThe police investigated the crime Agent PRED PatientArchitecture•Tested on the NomBank corpus (250,000 size)•[the crime’s ARG1] [investigation PRED] …•[the police’s ARG0] [investigation PRED] … •[The investigation PRED] of [the police ARG0?/ARG1?] …•Based on the current SOTA (Liu & Ng 2007)•Developed 12 new features:1) Syntactic Context: Agents are more likely to be in the sentence’s subject position:2) Animacy features: The most animate argument is more likely to be the Agent•Stanford Classifier (MaxEnt)Our contribution•We improved the current State of the Art results:0.60.650.70.750.80.850.90.95percent of training dataF1F1: ARG0,1F1: ARG2+L&N F1L&N F1: ARG0,1L&N F1: ARG2+Liu & Ng, 2007(Baseline)Us!0.40.450.50.550.60.650.70.750.80.851%3%5%10%20%30%40%50%60%70%80%90%100%percent of training dataF1F1: ARG0,1F1: ARG2+L&N F1L&N F1: ARG0,1L&N F1: ARG2+Our contribution•Especially over unseen predicate/constituents:Liu & Ng, 2007(Baseline)Us!Data analysis Syntactic position Animacy0%10%20%30%40%50%60%Sentence Subject Direct Object Other positionARG0ARG1ARG2+ARGM0%10%20%30%40%50%60%PER ORG LOCARG0ARG1ARG2+ARGMConclusions•Features modeling syntactic context and animacy improve nominal-Semantic Role Labeling•Consistently outperforms the current state of the art results:+.012 FB1 over all NomBank+.033 FB1 over unseen predicate/constituents•Greater improvements are
View Full Document