ConSmBop - Context-Dependent Bottom-up Semantic Parsing

Code for my ConSmBop project will be available soon.

• About
  • Acknowledgements

About

SmBoP is a semi-autoregressive bottom-up semantic parser, which takes an utterance and a DB schema as input, and constructs at decoding step $t$ the top-$K$ program sub-trees of height $≤ t$.

In this work I studied several methods to enhance SmBoP’s contextual modeling to improve performance when mapping a sequence of context-dependent user utterances to structured programs. I altered the encoder to address previous utterances, and enriched the decoder at each decoding step by reusing subtrees of the same height predicted in previous turns within the same interaction.

Acknowledgements

I thank Prof. Jonathan Berant and Ohad Rubin for their valuable guidance and feedback.