Domain adaptation in semantic role labeling using a neural language model and linguistic resources

摘要: We propose a method for adapting Semantic Role Labeling (SRL) systems from source domain to target by combining neural language model and linguistic resources generate additional training examples. primarily aim improve the results of Location, Time, Manner Direction roles. In our methodology, main words selected predicates arguments in source-domain data are replaced with domain. The replacement generated then filtered several filters (including Part-Of-Speech (POS), WordNet Predicate constraints). experiments on out-of-domain CoNLL 2009 data, Recurrent Neural Network Language Model (RNNLM) well-known semantic parser Lund University, we show enhanced recall F1 without penalizing precision four targeted These same SRL system using resources, better than that is trained examples enriched word embeddings. also demonstrate importance vocabulary when generating new