RoBox: CCG with Structured Perceptron for Supervised Semantic Parsing of Robotic Spatial Commands

Proceedings of the 8th International Workshop on Semantic Evaluation (SemEval 2014), pages 482–486, Dublin, Ireland, August 23-24, 2014. RoBox: CCG with Structured Perceptron for Supervised Semantic Parsing of Robotic Spatial Commands Kilian Evang University of Groningen k.evang@rug.nl Johan Bos University of Groningen johan.bos@rug.nl Abstract We use a Combinatory Categorial Gram- mar (CCG) parser with a structured per- ceptron learner to address Shared Task 6 of SemEval-2014, Supervised Semantic Parsing of Robotic Spatial Commands. Our system reaches an accuracy of 79% ignoring spatial context and 87% using the spatial planner, showing that CCG can successfully be applied to the task. 1 Introduction When interpreting utterances, humans use world knowledge whereas most semantic parsers to date rely purely on linguistic clues. Shared Task 6 in the SemEval 2014 campaign for semantic evalua- tion aims to integrate reasoning about microworlds with semantic parsing. In this task, a system is given an instruction for a robot and has to produce an executable semantic representation in Robot Control Language (Dukes, 2013a, RCL). The Robot Commands Treebank (Dukes, 2013b) is used for training and evaluation. We partici- pated in this shared task with a system rooted in Combinatory Categorial Grammar (CCG). In par- ticular, we were interested in finding out whether existing techniques for automatically deriving cat- egorial grammars with semantics could be moved easily to the new domain of robot commands and integrated with the provided spatial reasoning component. In this paper we outline our method and present the results for this shared task. 1 2 Extracting a CCG from RCL CCGs (Steedman, 2001) use a small set of atomic constituent categories such as S (sentence), NP This work is licensed under a Creative Commons At- tribution 4.0 International Licence. Page numbers and pro- ceedings footer are added by the organisers. Licence details: http://creativecommons.org/licenses/by/4.0/ 1Our code is available at http://www.let.rug. nl/evang/RoBox.zip (noun phrase) or PP (prepositional phrase). Con- stituents that take other constituents as arguments have complex categories describing their combi- natory potential. For example, an intransitive En- glish verb has category S\NP, meaning that it forms a sentence by combining with an NP to its left. Similarly, modifiers also have complex cate- gories. For example, a pre-sentential adverb might have category S/S because it combines with a sentence to its right to form a modified sentence. The combinatory rules that license these exam- ple combinations are called backward application and forward application. They and other combi- natory rules also allow for constituents to be asso- ciated with semantic expressions, and specify how to form a combined semantic expression for the derived larger constituent. In this section, we describe a process that takes an RCL corpus as input and produces a set of CCG lexical entries, i.e. natural-language words paired with categories and semantic expressions. The goal is for these lexical entries to produce the cor- rect semantics under CCG combinatory rules also for unseen robotic commands. 2.1 Transforming the Trees RCL expressions are rooted ordered trees whose nodes are labeled with tags. We will …

Content truncated to 500 words