Turku: Semantic Dependency Parsing as a Sequence Classification

Proceedings of the 9th International Workshop on Semantic Evaluation (SemEval 2015), pages 965–969, Denver, Colorado, June 4-5, 2015. c⃝2015 Association for Computational Linguistics Turku: Semantic Dependency Parsing as a Sequence Classification Jenna Kanerva1,2, Juhani Luotolahti1, Filip Ginter1 1 Department of Information Technology, University of Turku, Finland 2 University of Turku Graduate School (UTUGS), Turku, Finland jmnybl,mjluot,figint@utu.fi Abstract This paper presents the University of Turku entry to the SemEval-2015 task on Broad- Coverage Semantic Dependency Parsing. The system uses an existing transition-based parser as a sequence classifier to jointly predict all ar- guments of one candidate predicate at a time. Compared to our 2014 entry, the 2015 system gains about 3pp in terms of F-score for a frac- tion of the development time. Depending on the subtask, the difference between our entry and the winning system ranges between 1 and 5pp. 1 Introduction The SemEval-2015 task on Broad-Coverage Seman- tic Dependency Parsing is a continuation to the se- mantic parsing shared task organized for the first time in 2014. The objective of the shared task is to produce a rich semantic analysis for a given sen- tence in three distinct annotation formats. In contrast to the 2014 task, this year predicate disambiguation and two additional languages are included: Czech data from the Prague Czech–English Dependency Treebank (Hajiˇc et al., 2012) and Chinese data from the Penn Chinese Treebank (Xue et al., 2005). For English and Czech also out-of-domain test data is provided in order to test the generalization ability of the systems. The semantic parsing task includes three differ- ent tracks. In the closed track the systems must be trained using only the official training data, whereas in the open track all additional sources of infor- mation are allowed. Together with the training data the organizers provided also syntactic depen- dency parses produced in the Stanford Dependen- cies scheme (De Marneffe and Manning, 2008) with the dependency parser of Bohnet and Nivre (2012). In addition to the closed and open tracks, also a gold track is included, where gold standard dependency parses are given for both training and test data. This paper describes our system used to take part in the open and gold tracks of the shared task. The system is a sequence classifier built on top of an existing dependency parser. The main idea behind the implementation is to turn the task of predicting all arguments for a single predicate to a sequence classification problem, but still process each pred- icate independently. Predicting one predicate at a time feels very natural when working with data an- notated in PropBank style (Palmer et al., 2005), and since our main objective is to develop an SRL sys- tem optimized for Finnish PropBank (Haverinen et al., 2013), we did not want to merely follow the main methods from last year. Our system also re- quires syntactic analyses of the data, which is why we participated only on the tasks which allow their use (open and gold tracks). The system will be de- …

Content truncated to 500 words