Riga: from FrameNet to Semantic Frames with C6.0 Rules

Proceedings of the 9th International Workshop on Semantic Evaluation (SemEval 2015), pages 960–964, Denver, Colorado, June 4-5, 2015. c⃝2015 Association for Computational Linguistics Riga: from FrameNet to Semantic Frames with C6.0 Rules Guntis Barzdins, Peteris Paikens, Didzis Gosko University of Latvia, IMCS Rainis Blvd. 29, Riga, LV-1459, Latvia {guntis.barzdins,peteris.paikens,didzis.gosko}@lumii.lv Abstract For the purposes of SemEval-2015 Task-18 on the semantic dependency parsing we com- bined the best-performing closed track ap- proach from the SemEval-2014 competition with state-of-the-art techniques for FrameNet semantic parsing. In the closed track our sys- tem ranked third for the semantic graph accu- racy and first for exact labeled match of complete semantic graphs. These results can be attributed to the high accuracy of the C6.0 rule-based sense labeler adapted from the FrameNet parser. To handle large SemEval training data the C6.0 algorithm was extended to provide multi-class classification and to use fast greedy search without significant accura- cy loss compared to exhaustive search. A method for improved FrameNet parsing using semantic graphs is proposed. 1 Introduction The trend of natural language processing in recent years is shifting towards multilingual natural lan- guage understanding based on full-text shallow semantic parsing (e.g., Banarescu et al., 2013). De- spite various formalisms proposed, these ap- proaches are characterized by direct extraction of a bi-lexical semantic graph rather than a bi-lexical dependency tree from the surface form of the sen- tence. Following the best practice for semantic parsing established already by the SemEval-2014 Task 8 (Oepen et al., 2014) we modified the best- performing closed-track system there (Du et al., 2014) by removing some less essential components while adding a new component of our own. The newly added component is the C6.0 rule-based classifier (Barzdins et al., 2014) used both for graph parsing and for sense labeling. Sense label- ing is a novelty of SemEval-2015 Task 18 and was not present in the previous year competition. Se- mantic frame is comprised of a complete predica- tion combined with the sense identifier of its predicate as shown in Figure 1. Semantic frames are similar to FrameNet (Fillmore et al., 2003) frames, except that FrameNet argument labels are sense-specific – this mismatch can be resolved by feeding the semantic graph (instead of dependency tree) through the regular FrameNet parser. Figure 1. Semantic frame from the PSD corpus. We participated only in the closed track. Despite ranking third for the semantic graph accuracy, our system ranked first for exact labeled match of complete semantic graphs, and close second for semantic frame accuracy. 2 Baseline Architecture For semantic graph parsing we started by imple- menting a straight-forward baseline architecture described on the SemEval-2015 Task-18 evalua- tion page by the task organizers. The baseline ar- chitecture consists of two components: reduction 960 of the SDP graphs to trees and training the Mate- tools dependency parser (Bohnet, 2010) to produce such trees from the unparsed text. Instead of a de- structive reduction of the SDP graphs to trees, we implemented a fully reversible depth-first trans- formation from the …

Content truncated to 500 words