Universal Proposition Bank 2.0

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 1700–1711 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 1700 Universal Proposition Bank 2.0 Ishan Jindal1, Alexandre Rademaker1, Michał Ulewicz2∗, Linh Ha4, Huyen Nguyen3, Khoi-Nguyen Tran1, Huaiyu Zhu1, Yunyao Li4∗ 1IBM Research, 2Kyndryl, 3Vietnam National University, 4Apple ishan.jindal@ibm.com, alexrad@br.ibm.com, michal.ulewicz@kyndryl.com {huyenntm, hamylinh}@hus.edu.vn, kndtran@ibm.com, huaiyu@us.ibm.com, yunyaoli@apple.com Abstract Semantic role labeling (SRL) represents the meaning of a sentence in the form of predicate-argument structures. Such shallow semantic analysis is helpful in a wide range of downstream NLP tasks and real-world applications. As treebanks enabled the development of powerful syntactic parsers, high-quality training data in the form of propbanks is crucial to build models for accurate predicate-argument analysis. Unfortunately, most languages simply do not have corresponding propbanks due to the high cost required to construct such resources. To overcome such challenges, we released Universal Proposition Bank 1.0 (UP1.0) in 2017, with high-quality propbank data generated via a two-stage method exploiting monolingual SRL and multilingual parallel data. In this paper, we introduce Universal Proposition Bank 2.0 (UP2.0), with significant enhancements over UP1.0, including: (1) propbanks with higher quality by using a state-of-the-art monolingual SRL and improved auto-generation of annotations; (2) expanded language coverage (from 7 to 23 languages); (3) span annotation for the decoupling of syntactic analysis; and (4) gold data for a subset of the languages. We also share our experimental results that confirm the significant quality improvements of the generated propbanks. In addition, we present a comprehensive experimental evaluation on how different implementation choices impact the quality of the resulting data. We release these resources to the research community and hope to encourage more research on cross-lingual SRL. Keywords: Annotation projection, Semantic role labeling, Multilingual, Span-based SRL, Dependency-based SRL 1. Introduction Semantic role labeling (SRL) is a shallow semantic parsing task that identifies “who did what to whom when, where etc” for each predicate in a sentence. It provides an intermediate (shallow) level of a seman- tic representation that helps the map from syntactic parse structures to more fully-specified representations of meaning (Jurafsky and Martin, 2021). SRL has been shown to help a wide range of NLP applications such as natural language inference (Zhang et al., 2020b), ques- tion answering (Zhang et al., 2020b; Maqsud et al., 2014; Yih et al., 2016), machine translation (Shi et al., 2016) and information extraction (Niklaus et al., 2018; Zhang et al., 2020a). The increasing availability of manually annotated meaning representation datasets such as FrameNet (Fillmore et al., 2004), NomBank (Meyers et al., 2004), Proposition Bank (PropBank) (Palmer et al., 2005)) as well as significant advances in modeling techniques such as deep learning techniques have led to increased interest and progress in computational models for En- glish SRL. Parallelly, several attempts have been made to generate such resources for other languages such as German (Erk et al., 2003), Arabic (Zaghouani et al., 2010), Portuguese (Duran and Alu´ısio, 2011), Hindi (Vaidya et al., 2012), Finnish (Haverinen et al., 2015) …

Content truncated to 500 words