Abstract
Executable semantic parsing is the task of converting natural languageutterances into logical forms that can be directly used as queries to get aresponse. We build a transfer learning framework for executable semanticparsing. We show that the framework is effective for Question Answering (Q&A)as well as for Spoken Language Understanding (SLU). We further investigate thecase where a parser on a new domain can be learned by exploiting data on otherdomains, either via multi-task learning between the target domain and anauxiliary domain or via pre-training on the auxiliary domain and fine-tuning onthe target domain. With either flavor of transfer learning, we are able toimprove performance on most domains; we experiment with public data sets suchas Overnight and NLmaps as well as with commercial SLU data. We report thefirst parsing results on Overnight and state-of-the-art results on NLmaps. Theexperiments carried out on data sets that are different in nature show howexecutable semantic parsing can unify different areas of NLP such as Q&A andSLU.