Combining Semantic Annotation of Word Sense & Semantic Roles: A Novel Annotation Scheme for VerbNet Roles on German Language Data

Combining Semantic Annotation of Word Sense & Semantic Roles: A Novel Annotation Scheme for VerbNet Roles on German Language Data Éva Mújdricza-Maydt‡∗, Silvana Hartmann†, Iryna Gurevych†∗, Anette Frank‡∗ Department of Computational Linguistics, Heidelberg University‡ Ubiquitous Knowledge Processing Lab (UKP-TUDA), Department of Computer Science, Technische Universität Darmstadt† Research Training Group AIPHES, Heidelberg University and Technische Universität Darmstadt∗ {mujdricz | frank}@cl.uni-heidelberg.de, {gurevych | hartmann}@ukp.tu-darmstadt.de Abstract We present a VerbNet-based annotation scheme for semantic roles which we explore in an annotation study on German language data that combines word sense and semantic role annotation. We reannotate a substantial portion of the SALSA corpus with GermaNet senses and a revised scheme of VerbNet roles. We provide a detailed evaluation of the interaction between sense and role annotation. The resulting corpus will allow us to compare VerbNet role annotation for German to FrameNet and PropBank annotation by mapping to existing role annotations on the SALSA corpus. We publish the annotated corpus and detailed guidelines for the new role annotation scheme. Keywords: word sense annotation, semantic role annotation, GermaNet, VerbNet, SALSA corpus, guidelines, German 1. Introduction Semantic annotation of predicate-argument structure is an important task in NLP. During decades, different frameworks for representing semantic predicate-argument structure have been established, notably FrameNet, VerbNet and PropBank, with accompanying sense and role invento- ries and annotated resources, primarily for English (Baker et al., 1998; Kipper-Schuler, 2005; Palmer et al., 2005). The resources are interoperable via SemLink and the Uni- fied Verb Index (Loper et al., 2007). Combining GermaNet sense and VerbNet role annota- tion Within these established frameworks, the character- ization of predicates and roles differ in important ways: FrameNet defines semantic roles for verbs, nouns and ad- jectives evoking a frame. Due to its large frame inven- tory, and since roles are specific to a frame, there is a vast amount of roles to distinguish, and roles do not generalize across frames. PropBank defines six major roles (A0-A5); except for A0 and A1, their interpretation is not consis- tent across predicates, and role sets that characterize spe- cific senses are predicate-specific. In our work we focus on VerbNet. VerbNet operates with a closed set of seman- tic roles that are not predicate-specific, and hence can be applied to any predicate to characterize the semantic roles that arguments take in relation to that predicate. These differences in design have consequences for the scalability of automatic semantic role labeling (SRL) sys- tems, and the semantic expressiveness of the resulting an- notations. While FrameNet representations bear the most informative content, systems suffer from sparsity due to incomplete coverage. Frames also represent relatively coarse meanings, covering both synonyms and antonyms. PropBank has wide coverage and yields good automatic an- notation quality (Merlo and van der Plas, 2009), but the roles bear less semantic content. VerbNet-style role label- ing represents a middle ground between these frameworks, with a constrained role inventory of approximately 35 roles that offers interpretable role semantics and generalizability across predicates. In this work we focus on VerbNet-style role labeling …

Content truncated to 500 words