AT&T: The Tag&Parse Approach to Semantic Parsing of Robot Spatial Commands

Proceedings of the 8th International Workshop on Semantic Evaluation (SemEval 2014), pages 109–113, Dublin, Ireland, August 23-24, 2014. AT&T: The Tag&Parse Approach to Semantic Parsing of Robot Spatial Commands Svetlana Stoyanchev, Hyuckchul Jung, John Chen, Srinivas Bangalore AT&T Labs Research 1 AT&T Way Bedminster NJ 07921 {sveta,hjung,jchen,srini}@research.att.com Abstract The Tag&Parse approach to semantic parsing first assigns semantic tags to each word in a sentence and then parses the tag sequence into a semantic tree. We use statistical approach for tagging, pars- ing, and reference resolution stages. Each stage produces multiple hypotheses which are re-ranked using spatial validation. We evaluate the Tag&Parse approach on a cor- pus of Robotic Spatial Commands as part of the SemEval Task6 exercise. Our sys- tem accuracy is 87.35% and 60.84% with and without spatial validation. 1 Introduction In this paper we describe a system participating in the SemEval2014 Task-6 on Supervised Seman- tic Parsing of Robotic Spatial Commands. It pro- duces a semantic parse of natural language com- mands addressed to a robot arm designed to move objects on a grid surface. Each command directs a robot to change position of an object given a current configuration. A command uniquely iden- tifies an object and its destination, for example “Move the turquoise pyramid above the yellow cube”. System output is a Robot Control Lan- guage (RCL) parse (see Figure 1) which is pro- cessed by the robot arm simulator. The Robot Spa- tial Commands dataset (Dukes, 2013) is used for training and testing. Our system uses a Tag&Parse approach which separates semantic tagging and semantic parsing stages. It has four components: 1) semantic tag- ging, 2) parsing, 3) reference resolution, and 4) spatial validation. The first three are trained using LLAMA (Haffner, 2006), a supervised machine learning toolkit, on the RCL-parsed sentences. This work is licensed under a Creative Commons Attribution 4.0 International Licence. Page numbers and proceedings footer are added by the organisers. Licence details: http: //creativecommons.org/licenses/by/4.0/ For semantic tagging, we train a maximum en- tropy sequence tagger for assigning a semantic la- bel and value to each word in a sentence, such as type cube or color blue. For parsing, we train a constituency parser on non-lexical RCL semantic trees. For reference resolution, we train a maxi- mum entropy model that identifies entities for ref- erence tags found by previous components. All of these components can generate multiple hypothe- ses. Spatial validation re-ranks these hypotheses by validating them against the input spatial con- figuration. The top hypothesis after re-ranking is returned by the system. Separating tagging and parsing stages has sev- eral advantages. A tagging stage allows the system flexibility to abstract from possible grammatical or spelling errors in a command. It assigns a seman- tic category to each word in a sentence. Words not contributing to the semantic meaning are assigned ‘O’ label by the tagger and are ignored in the fur- ther processing. Words that are misspelled can po- tentially receive a correct tag when a word simi- larity feature is …

Content truncated to 500 words