ShrdLite: Semantic Parsing Using a Handmade Grammar

Proceedings of the 8th International Workshop on Semantic Evaluation (SemEval 2014), pages 556–559, Dublin, Ireland, August 23-24, 2014. ShrdLite: Semantic Parsing Using a Handmade Grammar Peter Ljunglöf Department of Computer Science and Engineering University of Gothenburg and Chalmers University of Technology Gothenburg, Sweden peter.ljunglof@cse.gu.se Abstract This paper describes my approach for parsing robot commands, which was task 6 at SemEval 2014. My solution is to manually create a compact unifica- tion grammar. The grammar is highly am- biguous, and relies heavily on filtering the parse results by checking their consistency with the current world. The grammar is small, consisting of not more than 25 grammatical and 60 lexical rules. The parser uses simple error correc- tion together with a straightforward itera- tive deepening search. Nevertheless, with these very basic algorithms, the system still managed to get 86.1% correctness on the evaluation data. Even more interesting is that by making the parser slightly more robust, the accuracy of the system rises to 93.5%, and by adding one single word to the lexicon, the accuracy is boosted to 98.0%. 1 Introduction SemEval 2014, task 6, was about parsing com- mands to a robot operating in a blocks world. The goal is to parse utterances in natural language into commands in a formal language, the Robot Con- trol Language (RCL). As a guide the system can use a spatial planner which can tell whether an RCL command is meaningful in a given blocks world. The utterances are taken from the Robot Com- mands Treebank (Dukes, 2013), which pairs 3409 sentences with semantic annotations consisting of an RCL command together with a description of This work is licensed under a Creative Commons Attribu- tion 4.0 International Licence. Page numbers and proceed- ings footer are added by the organisers. Licence details: http://creativecommons.org/licenses/by/4.0/ a world where the command is meaningful. The corpus was divided into 2500 training sentences and 909 evaluation sentences. The system that is described in this paper, to- gether with the evaluation data, is available online from GitHub: https://github.com/heatherleaf/semeval-2014-task6 2 System Description The Shrdlite system is based on a small unification grammar, together with a naive robust parser im- plemented using iterative deepening. After pars- ing, the resulting parse trees are modified accord- ing to six extra-grammatical semantic rules. The grammar and the semantic rules were hand- crafted using manual analysis of the available 2500 training sentences, and an incremental and iterative process to select and fine-tune the gram- mar. The total amount of work for creating the grammar consisted of about 3–4 days for one per- son. This excludes programming the robust parser and the rest of the system, which took another 2–3 days. I did not have any access to the 909 evalua- tion sentences while developing the grammar or the other parts of the system. 2.1 Grammar The grammar is a Prolog DCG unification gram- mar (Pereira and Warren, 1980) which builds a se- mantic parse tree during parsing. The core gram- mar is shown in figure 1. For presentation …

Content truncated to 500 words