Copenhagen-Malmö: Tree Approximations of Semantic Parsing Problems

Proceedings of the 8th International Workshop on Semantic Evaluation (SemEval 2014), pages 213–217, Dublin, Ireland, August 23-24, 2014. Copenhagen-Malm¨o: Tree Approximations of Semantic Parsing Problems Natalie Schluter†, Jakob Elming, Sigrid Klerke, H´ector Mart´ınez Alonso, Dirk Hovy Barbara Plank, Anders Johannsen, and Anders Søgaard †Dpt. of Computer Science Center for Language Technology Malm¨o University University of Copenhagen natalie.schluter@mah.se {zmk867,skl,alonso}@hum.ku.dk {dirk,bplank}@cst.dk, {ajohannsen,soegaard}@hum.ku.dk Abstract In this shared task paper for SemEval- 2014 Task 8, we show that most se- mantic structures can be approximated by trees through a series of almost bijective graph transformations. We transform in- put graphs, apply off-the-shelf methods from syntactic parsing on the resulting trees, and retrieve output graphs. Us- ing tree approximations, we obtain good results across three semantic formalisms, with a 15.9% error reduction over a state- of-the-art semantic role labeling system on development data. Our system came in 3/6 in the shared task closed track. 1 Introduction Semantic analyses often go beyond tree- structured representations, assigning multiple se- mantic heads to nodes, some semantic formalisms even tolerating directed cycles.1 At the same time, syntactic parsing is a mature field with effi- cient, highly optimised decoding and learning al- gorithms for tree-structured representations. We present tree approximation algorithms that in com- bination with a state-of-the-art syntactic parser achieve competitive performance in semantic di- graph parsing. We investigate two kinds of tree approximation algorithms that we will refer to as pruning algo- rithms and packing algorithms. Our pruning al- gorithms simply remove and reverse edges until the graph is a tree; edge reversals are then undone as a postprocessing step. Our packing algorithms, on the other hand, carry out two bijective graph This work is licenced under a Creative Commons Attribu- tion 4.0 International License. Page numbers and proceed- ings footer are added by the organizers. http://creativecommons.org/licenses/by/4.0/License details: 1For example, HPSG predicate-argument structures (Pol- lard and Sag, 1994). transformations to pack structural information into new edge labels, making it possible to reconstruct most of the structural complexity as a postprocess- ing step. Specifically, we present a packing al- gorithm that consists of two fully bijective graph transformations, in addition to a further transfor- mation that incurs only a small information loss. We carry out experiments across three seman- tic annotations of the Wall Street Journal section of the Penn Treebank (Marcus et al., 1993), cor- responding to simplified versions of the semantic formalisms minimal recursion semantics (MRS) (Copestake et al., 2005), Enju-style predicate- argument structures (Miyao and Tsujii, 2003), and Prague-style tectogrammar semantics (B¨ohmov´a et al., 2003). We show that pruning and pack- ing algorithms lead to state-of-the-art performance across these semantic formalisms using an off-the- shelf syntactic dependency parser. 2 Related work Sagae and Tsujii (2008) present a pruning algo- rithm in their paper on transition-based parsing of directed acyclic graphs (DAGs), which discards the edges of longest span entering nodes. They apply the dependency parser described in Sagae and Tsujii (2007) to the tree representations. We note that this algorithm is not sufficient to produce trees …

Content truncated to 500 words