Semantic Role Labeling Frameworks -- a low-resource task with straightforward cases and open challenges

There are three main, well known semantic role labeling (SRL) frameworks: ProbBank, VerbNet, and FrameNet [1]. 
However, they are not utilized equally, nor are they deployed in the same way. The community is aware their theoretical and application-sensitive differences between these frameworks, while all of them intend to be a semantic "sense-giving" of predicate argument structures, thus interpreting events and participants in context with semantically tactile labels.

We make the three frameworks comparable with a small German parallel dataset [2], which can serve as a basis for further empirical investigations. The presentation will describe 1) the data set, 2) its corpus building process, and 3) first experiments on automatic SRL trained on the dataset.

In particular, we present
1) the corpus building process
  - with our findings on best practices elaborated during the annotation study
  - the developed VerbNet-style annotation framework with GermaNet predicate senses and our adapted and precisely described VerbNet roles
2) the parallel dataset
  - including a comparison of the annotations from the three frameworks 
3) automatic SRL experiments on the dataset
  - trained with mateplus [3]
  - as well as training data expansion results
  - which give an outline for further plans 
  
[1] FrameNet: Charles J Fillmore, Christopher R Johnson, and Miriam RL Petruck (2003). Background to FrameNet. International journal of lexicography, 16(3):235–250.; 
    PropBank: Martha Palmer, Daniel Gildea, and Paul Kingsbury. 2005. The Proposition Bank: An annotated corpus of semantic roles. Computational linguistics, 31(1): 71–106.; 
    VerbNet: Karin Kipper Schuler. 2005. Verbnet: A broad-coverage, comprehensive verb lexicon.
[2] Éva Mújdricza-Maydt, Silvana Hartmann, Iryna Gurevych, and Anette Frank (2016). Combining Semantic Annotation of Word Sense & Semantic Roles: A Novel Annotation Scheme for VerbNet Roles on German Language Data. In Proceedings of the Tenth International Conference on Language Resources and Evaluation (LREC 2016), 3031–3038, Paris, France.
[3] Michael Roth and Kristian Woodsend (2014). Composition of word representations improves semantic role labelling. Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), Doha, Qatar, October, pp. 407-413.