- Abstract: Semantic parsing which maps a natural language sentence into a formal machine-readable representation of its meaning, is highly constrained by the limited annotated training data. Inspired by the idea of coarse-to-fine, we propose a general-to-detailed neural network(GDNN) by incorporating cross-domain sketch(CDS) among utterances and their logic forms. For utterances in different domains, the General Network will extract CDS using an encoder-decoder model in a multi-task learning setup. Then for some utterances in a specific domain, the Detailed Network will generate the detailed target parts using sequence-to-sequence architecture with advanced attention to both utterance and generated CDS. Our experiments show that compared to direct multi-task learning, CDS has improved the performance in semantic parsing task which converts users' requests into meaning representation language(MRL). We also use experiments to illustrate that CDS works by adding some constraints to the target decoding process, which further proves the effectiveness and rationality of CDS.
- Keywords: semantic parsing, natural language understanding, machine learning
- TL;DR: General-to-detailed neural network(GDNN) with Multi-Task Learning by incorporating cross-domain sketch(CDS) for semantic parsing