Understanding Instructions on Large Scale for Human-Robot Interaction

摘要

Correctly interpreting human instructions is the first step to human-robot interaction. Previous approaches to semantically parsing the instructions relied on large numbers of training examples with annotation to widely cover all words in a domain. Annotating large enough instructions with semantic forms needs exhaustive engineering efforts. Hence, we propose propagating the semantic lexicon to learn a semantic parser from limited annotations, whereas the parser still has the ability of interpreting instructions on a large scale. We assume that the semantically-close words have the same semantic form based on the fact that human usually uses different words to refer to a same object or task. Our approach softly maps the unobserved words/phrases to the semantic forms learned from the annotated copurs through a metric for knowledge-based lexical similarity. Experiments on the collected instructions showed that the semantic parser learned with lexicon propagation outperformed the baseline. Our approach provides an opportunity for the robots to understand the human instructions on a large scale.