QANom: Question-Answer driven SRL for Nominalizations

Proceedings of the 28th International Conference on Computational Linguistics, pages 3069–3083 Barcelona, Spain (Online), December 8-13, 2020 3069 QANom: Question-Answer driven SRL for Nominalizations Ayal Klein1 Jonathan Mamou2 Valentina Pyatkin1 Daniela Brook Weiss1 Hangfeng He3 Dan Roth3 Luke Zettlemoyer4 Ido Dagan1 1Computer Science Department, Bar Ilan University 2Intel Labs, Israel 3University of Pennsylvania 4University of Washington {ayal.s.klein,jonathan.mamou,valpyatkin, daniela.stepanov}@gmail.com {hangfeng,danroth}@seas.upenn.edu lsz@cs.washington.edu dagan@cs.biu.ac.il Abstract We propose a new semantic scheme for capturing predicate-argument relations for nominaliza- tions, termed QANom. This scheme extends the QA-SRL formalism (He et al., 2015), modeling the relations between nominalizations and their arguments via natural language question-answer pairs. We construct the first QANom dataset using controlled crowdsourcing, analyze its quality and compare it to expertly annotated nominal-SRL annotations, as well as to other QA-driven annotations. In addition, we train a baseline QANom parser for identifying nominalizations and labeling their arguments with question-answer pairs. Finally, we demonstrate the extrinsic utility of our annotations for downstream tasks using both indirect supervision and zero-shot settings. 1 Introduction Semantic Role Labeling (SRL) is the prominent representation for annotating predicate-argument struc- tures. SRL annotations were shown useful for various downstream tasks, such as machine comprehension (Wang et al., 2015), cross-document coreference (Barhom et al., 2019), dialog (Chen et al., 2013) and summarization (Trandab˘at¸, 2011). Traditionally, SRL research is biased toward focusing on verbal predi- cates, as evident by their dominance in large scale semantic resources, such as PropBank (Kingsbury and Palmer, 2002), FrameNet (Baker et al., 1998) and OntoNotes (Pradhan et al., 2013), and consequentially among SRL models (He et al., 2018; Tan et al., 2018; Strubell et al., 2018). Nevertheless, other types of predicates, such as nominalizations, are frequent in natural language, which also draw some research attention (Hajic et al., 2009; Jiang and Ng, 2006; Zhao and Titov, 2020). As a significant milestone, the NomBank initiative (Meyers et al., 2004a) provided extensive predicate- argument annotations for various types of nouns. In particular, since deverbal nominalizations share an underlying argument structure with their verbal counterparts, NomBank annotates these by applying the same annotation scheme as PropBank. A verb-derived noun will be mapped to a frame file shared by PropBank’s verbal predicates, and accordingly a shared role-set. This design principle is meant for converging the semantic role representation of deverbal nominalizations with their corresponding verbal predicates, thus abstracting semantic content over surface realization specifics (Meyers et al., 2004b). Annotating SRL resources involves substantial effort and cost. This hinders research progress, as it is hard to extend large-scale resources to additional text genres and languages. In order to address this data collection barrier, question-answer driven semantic role labeling (QA-SRL) (He et al., 2015) was proposed as a natural, easily attainable formulation of SRL. QA-SRL labels each predicate-argument relationship with a question-answer pair, where natural language questions represent semantic roles, and answers correspond to arguments (see Table 1 for a comparative illustration). This format yields rich and This work is licensed under a Creative Commons Attribution 4.0 International License. License details: http:// creativecommons.org/licenses/by/4.0/. 3070 Thomas …

Content truncated to 500 words