Jump to Content

Answering Conversational Questions on Structured Data without Logical Forms

Thomas Müller
Francesco Piccinno
Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing, Association for Computational Linguistics (2019)


We present a novel approach to answering sequential questions based on structured objects such as knowledge bases or tables without using a logical form as an intermediate representation. We encode tables as graphs using a graph neural network model based on the Transformer architecture. The answers are then selected from the encoded graph using a pointer network. This model is appropriate for processing conversations around structured data, where the attention mechanism that selects the answer to a question can also be used to resolve conversational references. We demonstrate the validity of this approach with competitive results on the Sequential Question Answering task (SQA) (Iyyer et al., 2017).