Turku: Broad-Coverage Semantic Parsing with Rich Features

Proceedings of the 8th International Workshop on Semantic Evaluation (SemEval 2014), pages 678–682, Dublin, Ireland, August 23-24, 2014. Turku: Broad-Coverage Semantic Parsing with Rich Features Jenna Kanerva∗ Department of Information Technology University of Turku Finland jmnybl@utu.fi Juhani Luotolahti∗ Department of Information Technology University of Turku Finland mjluot@utu.fi Filip Ginter Department of Information Technology University of Turku Finland figint@utu.fi Abstract In this paper we introduce our system ca- pable of producing semantic parses of sen- tences using three different annotation for- mats. The system was used to partic- ipate in the SemEval-2014 Shared Task on broad-coverage semantic dependency parsing and it was ranked third with an overall F1-score of 80.49%. The sys- tem has a pipeline architecture, consisting of three separate supervised classification steps. 1 Introduction In the SemEval-2014 Task 8 on semantic parsing, the objective is to extract for each sentence a rich set of typed semantic dependencies in three differ- ent formats: DM, PAS and PCEDT. These formats differ substantially both in the assignment of se- mantic heads as well as in the lexicon of seman- tic dependency types. In the open track of the shared task, participants were encouraged to use all resources and tools also beyond the provided training data. To improve the comparability of the systems, the organizers provided ready-to-use de- pendency parses produced using the state-of-the- art parser of Bohnet and Nivre (2012). In this paper we describe our entry in the open track of the shared task. Our system is a pipeline of three support vector machine classifiers trained separately for detecting semantic dependencies, assigning their roles, and selecting the top nodes of semantic graphs. In this, we loosely follow the architecture of e.g. the TEES (Bj¨orne et al., 2012) and EventMine (Miwa et al., 2012) systems, which were found to be effective in the structurally ∗These authors contributed equally. This work is licenced under a Creative Commons Attribution 4.0 International License. Page numbers and proceedings footer are added by the organizers. License details: http: //creativecommons.org/licenses/by/4.0/ related task of biomedical event extraction. Sim- ilar classification approach is shown to be effec- tive also in semantic parsing by e.g. Zhao et al. (2009), the winner of the CoNLL’09 Shared Task on Syntactic and Semantic Dependencies in Mul- tiple Languages (SRL-only subtask) (Hajiˇc et al., 2009), where semantic parsing is approached as a word-pair classification problem and semantic ar- guments and their roles are predicted simultane- ously. In preliminary experiments, we also de- veloped a joint approach to simultaneously iden- tify semantic dependencies and assign their roles, but found that the performance of the joint predic- tion was substantially worse than for the current pipeline approach. As the source of features, we rely heavily on the syntactic parses as well as other external resources such as vector space represen- tations of words and large-scale syntactic n-gram statistics. In the following sections, we describe the three individual classification steps of our semantic parsing pipeline. 2 Detecting Semantic Dependencies The first step of our semantic parsing pipeline is …

Content truncated to 500 words