Automatic Propbank Generation for Turkish

Proceedings of Recent Advances in Natural Language Processing, pages 33–41, Varna, Bulgaria, Sep 2–4, 2019. https://doi.org/10.26615/978-954-452-056-4_005 33 Automatic Propbank Generation for Turkish Koray Ak Department of Computer Engineering Is¸ık University ˙Istanbul, Turkey koray.ak@isik.edu.tr Olcay Taner Yıldız Department of Computer Engineering Is¸ık University ˙Istanbul, Turkey olcaytaner@isikun.edu.tr Abstract Semantic role labeling (SRL) is an im- portant task for understanding natural lan- guages, where the objective is to analyse propositions expressed by the verb and to identify each word that bears a semantic role. It provides an extensive dataset to en- hance NLP applications such as informa- tion retrieval, machine translation, infor- mation extraction, and question answer- ing. However, creating SRL models are difficult. Even in some languages, it is in- feasible to create SRL models that have predicate-argument structure due to lack of linguistic resources. In this paper, we present our method to create an automatic Turkish PropBank by exploiting parallel data from the translated sentences of En- glish PropBank. Experiments show that our method gives promising results. 1 Introduction Semantic role labeling (SRL) is a well defined task that identifies semantic roles of the words in a sentence. Event characteristics and partici- pants are simply identified by answering “Who did what to whom” questions. Having this semantic information facilitates NLP applications such as machine translation, information extraction, and question answering. After the development of sta- tistical machine learning methods in the area of computational linguistics, learning complex lin- guistic knowledge has became feasible for NLP applications. Recent semantic resources specifi- cally for SRL which provides input for develop- ing statistical approaches are FrameNet (Fillmore et al., 2004), PropBank (Kingsbury and Palmer, 2002) (2003), (2005), (Bonial et al., 2014) and NomBank (2004). These resources enables us to understand language structure by providing a sta- ble semantic representation. Among these resources PropBank is a com- monly used semantic resource which includes predicate - argument structure by stating the roles that each predicate can take along with the anno- tated corpora. It has been applied to more than 15 different languages. However, manually creating such semantic resource is labor-intensive, time- consuming and most importantly requires a pro- fessional linguistic perspective. Also limited lin- guistic data further blocks generating PropBank- like resources. Various studies such as Zhuang and Zong (2010), Van der Plas et al. (2011) (2014), Kozhevnikov and Titov (2013), Akbik et al. (2015), which transfer semantic information using parallel corpus, are presented to cope with these problems. In this way, semantic information pro- jected from a resource-rich language (English) to a language with inadequate resources and Prop- Bank of the target language is automatically gen- erated. Here the assumption is translated parallel sentences generally share same semantic informa- tion. Word and constituent based alignment tech- niques are widely used to construct mapping be- tween source and target languages for annotation projection. Previous studies report translation di- vergences and language specific differences affect the quality of the projection. Filtering projections using learning methods is suggested to increase precision. In this paper, we present our …

Content truncated to 500 words