Semantic Role Labeling as Dependency Parsing: Exploring Latent Tree Structures inside Arguments

Proceedings of the 29th International Conference on Computational Linguistics, pages 4212–4227 October 12–17, 2022. 4212 Semantic Role Labeling as Dependency Parsing: Exploring Latent Tree Structures Inside Arguments Yu Zhang⋆, Qingrong Xia⋆�, Shilin Zhou⋆, Yong Jiang☼, Guohong Fu⋆˚, Min Zhang⋆ ⋆Institute of Artificial Intelligence, School of Computer Science and Technology, Soochow University, Suzhou, China �Huawei Cloud, China ☼DAMO Academy, Alibaba Group, China {yzhang.cs,slzhou.cs}@outlook.com; xiaqingrong@huawei.com yongjiang.jy@alibaba-inc.com; {ghfu,minzhang}@suda.edu.cn Abstract Semantic role labeling (SRL) is a fundamen- tal yet challenging task in the NLP community. Recent works of SRL mainly fall into two lines: 1) BIO-based; 2) span-based. Despite ubiq- uity, they share some intrinsic drawbacks of not considering internal argument structures, po- tentially hindering the model’s expressiveness. The key challenge is arguments are flat struc- tures, and there are no determined subtree real- izations for words inside arguments. To remedy this, in this paper, we propose to regard flat ar- gument spans as latent subtrees, accordingly re- ducing SRL to a tree parsing task. In particular, we equip our formulation with a novel span- constrained TreeCRF to make tree structures span-aware and further extend it to the second- order case. We conduct extensive experiments on CoNLL05 and CoNLL12 benchmarks. Re- sults reveal that our methods perform favorably better than all previous syntax-agnostic works, achieving new state-of-the-art under both end- to-end and w/ gold predicates settings. 1 Introduction Semantic role labeling (SRL) is a fundamental yet challenging task in the NLP community, involving predicate and argument identification, as well as semantic role classification. As SRL can provide informative linguistic representations, it has been widely adopted in downstream tasks like question answering (Berant et al., 2013; Yih et al., 2016), information extraction (Christensen et al., 2010; Lin et al., 2017), and machine translation (Liu and Gildea, 2010; Bazrafshan and Gildea, 2013), etc. Recent works of SRL mainly fall into two lines: 1) BIO-based; 2) span-based. The former views SRL as a sequence labeling task (Zhou and Xu, 2015; Strubell et al., 2018; Shi and Lin, 2019). For each predicate, each token is tagged with a label starting with BIO prefixes indicating if it is at the Beginning, Inside, or Outside of an argument. ˚Corresponding author ... in other European markets ... closed ... NULL NULL NULL NULL AM-LOC Figure 1: An argument example (below) and its related subtree structure (above) for the predicate “closed”. The latter (He et al., 2018a; Ouchi et al., 2018; Li et al., 2019), in contrast, opts to jointly predict all predicate and argument span pairs using a span- graph formulation. Despite ubiquity, there are some drawbacks that limit the expressiveness of the two methods. First, framing predicate-argument structures as a BIO- tagging scheme is less effective as it lacks explicit modeling of span-level representations, so that long adjacencies of argument phrases can be ignored (Cohn and Blunsom, 2005; Jie and Lu, 2019; Zhou et al., 2020d; Xu et al., 2021). Second, span-based method seeks to pick very few (typically ă10%) positive examples from Opn3q candidate predicate- argument pairs, thus suffering …

Content truncated to 500 words