Abstract

Research on question answering with knowledge base has recently seen anincreasing use of deep architectures. In this extended abstract, we study theapplication of the neural machine translation paradigm for question parsing. Weemploy a sequence-to-sequence model to learn graph patterns in the SPARQL graphquery language and their compositions. Instead of inducing the programs throughquestion-answer pairs, we expect a semi-supervised approach, where alignmentsbetween questions and queries are built through templates. We argue that thecoverage of language utterances can be expanded using late notable works innatural language generation.

 

Quick Read (beta)