- Theory neutral numeric core roles (Arg0, Arg1, etc.)
- Interpretation of roles: verb-specific framesets
- Arg0 and Arg1 usually correspond to prototypical Agent and Patient/Theme roles.
- Other arguments do not consistently generalize across verbs
- Though there is an effort to make roles consistent. Arg2 is usually instrument, benefactive, attribute; Arg3 usually corresponds to starting point, benefactive, attribute and ARG4 is usually ending point.
- Different senses have different framesets
- Syntactic alternations that preserve meaning are kept toghether in a single frameset
- Closed set of 13 general labels for Adjuncts (e.g., Temporal, Manner, Location, etc.)
Failure to handle multi-word expressions: "The 12K rolesets in PB describe mostly single word predicates, to a great part leaving aside multiword expressions (MWEs). Complex predicates (CPs), ‘predicates which are multi-headed: they are composed of more than one grammatical element’ (Ramisch, 2012), are most relevant in the context of SRL. Light verb constructions (LVCs), e.g. take care, and verb particle constructions (VPCs), e.g. watch out, are the most frequently occurring types of CPs. As Bonial et al. (2014) stated ‘PB has previously treated language as if it were purely compositional, and has therefore lumped the majority of MWEs in with lexical verb usages’. For example the predicates in the CPs take a hard line, take time and many others are all annotated with a sense of take, meaning acquire, come to have, chose, bring with you from somewhere. This results in a loss of semantic information in the PB annotations.
This is especially critical because CPs are a frequent phenomenon. The Wiki50 corpus (Vincze et al., 2011), which provides a full coverage MWE annotation, counts 814 occurrences of LVCs and VPCs in 4350 sentences. This makes for one CP in every fifth sentence." (Jagfeld and van der Plas, 2015)
Languages other than English Edit
- Chinese PropBank
- Korean PropBank
- AnCora corpus: Spanish and Catalan
- Prague Dependency Treebank: Czech
- Penn Arabic TreeBank: Arabic
- Turkish PropBank
- ↑ Jagfeld, Glorianna, and Lonneke van der Plas. "Towards a Better Semantic Role Labeling of Complex Predicates." NAACL-HLT 2015 Student Research Workshop (SRW). 2015.
- ↑ Hwang J, Bhatia A, Bonial C, Mansouri A, Vaidya A, Xue N, Palmer M. 2010. PropBank Annotation of Multilingual Light Verb Constructions
- ↑ Bonial, Claire, Julia Bonn, Kathryn Conger, Jena D. Hwang, and Martha Palmer. "PropBank: Semantics of New Predicate Types." In LREC, pp. 3013-3019. 2014.
- This article is in an early stage. Action is needed to make it more useful.