An Open Vocabulary Semantic Parser for End-User Programming Using Natural Language

The ability to automatically interpret natural language commands and actions has the potential of freeing up endusers to interact with software artefacts without the syntactic, vocabulary and formal constraints of a programming language. As most semantic parsers for end-user programming have been operating under a restricted vocabulary setting, it is unclear how these approaches perform over conditions of high semantic heterogeneity (e.g. in an open vocabulary). As the generation of annotated data is costly and time-consuming, models that effectively address complex learning problems constrained under the assumption of small annotated data sets are highly relevant. In this paper, we propose a semantic parsing approach to map natural language commands to actions from a large and heterogeneous frame set trained under a small set of annotated data. The semantic parsing approach uses the combination of semantic role labelling, distributional semantics geometric features and semantic pivoting in order to address the semantic matching problem in an open vocabulary setting.