Sieve-Based Spatial Relation Extraction with Expanding Parse Trees

Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 758–768, Lisbon, Portugal, 17-21 September 2015. c⃝2015 Association for Computational Linguistics. Sieve-Based Spatial Relation Extraction with Expanding Parse Trees Jennifer D’Souza and Vincent Ng Human Language Technology Research Institute University of Texas at Dallas Richardson, TX 75083-0688 {jld082000,vince}@hlt.utdallas.edu Abstract A key challenge introduced by the re- cent SpaceEval shared task on spatial re- lation extraction is the identification of MOVELINKs, a type of spatial relation in which up to eight spatial elements can participate. To handle the complexity of extracting MOVELINKs, we combine two ideas that have been successfully applied to information extraction tasks, namely tree kernels and multi-pass sieves, propos- ing the use of an expanding parse tree as a novel structured feature for train- ing MOVELINK classifiers. Our approach yields state-of-the-art results on two key tasks in SpaceEval. 1 Introduction Spatial relation extraction is the task of determin- ing the relation among a set of spatial elements. Although it has thus far received much less atten- tion than temporal relation extraction, there has been a surge of interest in it in recent years, as evidenced by the organization of the three shared tasks on spatial relation extraction, namely the spatial role labeling tasks in 2012 (Kordjamshidi et al., 2012) and 2013 (Kolomiyets et al., 2013), as well as this year’s SpaceEval task (Pustejovsky et al., 2015). The task has also evolved over the years, with new types of spatial elements and/or spatial relations being defined in each shared task. For instance, while the first two shared tasks have focused on extracting spatial relations between stationary objects, SpaceEval examines for the first time spatial relations on objects in motion. Extracting spatial relations on objects in mo- tion, or MOVELINKs, is very challenging. The challenge stems in part from the fact that a MOVELINK involves two mandatory participants (with roles mover and trigger) and up to six op- tional participants (with other semantic roles). As Figure 1: A MOVELINK example. an example, consider the MOVELINK that can be extracted from the sentence “John walked from Boston to Cambridge”. As shown in Figure 1, this MOVELINK involves six spatial elements: “John” as the mover, “walked” as the trigger, “Boston” as the source, “Cambridge” as the goal, and “from” and “to” as the motion signals. Given the complexity of MOVELINKs, any ap- proach that attempts to jointly identify all the spa- tial elements involved in a MOVELINK and their roles is computationally infeasible. On the other extreme, one can tackle the task by identifying each element involved in a MOVELINK indepen- dently of the other elements. In fact, this is roughly the approach adopted by our participating sys- tem in SpaceEval (D’Souza and Ng, 2015), which achieved the best results in one of the SpaceE- val tasks involving MOVELINK extraction. Specif- ically, this system trains one classifier for each optional role r to identify the filler for r in a MOVELINK independently of the other optional roles. Although this …

Content truncated to 500 words