Extracting Relations between Non-Standard Entities using Distant Supervision and Imitation Learning

Distantly supervised approaches have be-come popular in recent years as they allowtraining relation extractors without text-bound annotation, using instead knownrelations from a knowledge base and alarge textual corpus from an appropri-ate domain. While state of the art dis-tant supervision approaches use off-the-shelf named entity recognition and clas-sification (NERC) systems to identify re-lation arguments, discrepancies in domainor genre between the data used for NERCtraining and the intended domain for therelation extractor can lead to low perfor-mance. This is particularly problematicfor “non-standard” named entities such asalbumwhich would fall into theMISCcategory. We propose to ameliorate thisissue by jointly training the named entityclassifier and the relation extractor usingimitation learning which reduces struc-tured prediction learning to classificationlearning. We further experiment withWeb features different features and com-pare against using two off-the-shelf su-pervised NERC systems, Stanford NERand FIGER, for named entity classifica-tion. Our experiments show that imita-tion learning improves average precisionby 4 points over an one-stage classificationmodel, while removing Web features re-sults in a 6 points reduction. Compared tousing FIGER and Stanford NER, averageprecision is 10 points and 19 points higherwith our imitation learning approach.