Lexicon Propagation for Learning a Large-scale Semantic Parser

For the purpose of smooth human-robot interaction, a robot is supposed to be capable of semantically parsing the human instructions in a large scale. However, the existing supervised approaches to learning a large-scale semantic parser needs a good deal of training examples with annotations. The exhaustive cost of annotating enough sentences prevents them from learning such parser for interpreting instructions. One of the reasons is that a small number of training examples result in a parser with the lexcion having low coverage on words/phrases of a domain. Hence, we introduce a semi-supervised approach to propagating lexicon based on the assumption that similar words have similar semnatic forms. Our approach first learns a seed lexicon from annotated corpus then smoothly maps unobserved words/phrases into those having already learned. Experiments on instructions, which were collected for the tasks in domestic environment, shows that our semantic parser with lexicon propagation improves by 30.28% F1-measure over the one learned via purely supervised algorithm.