Conversational semantic parsing over tables by decoupling and grouping actions

Abstract In this article, we present a new approach for answering a series of simple but inter-related questions over a table by mapping them to the corresponding SQLs. We follow the sequence-to-action paradigm and contribute two innovations: (1) an improved action space. We decouple some actions from existing work because they were jointly deciding unrelated query components, and group some actions with one neural network feature extractor as they were seeking the same aspect of information from the question and table. We also extend the equal action to match multiple cells under a column with a key word or phrase. (2) a simplified scoring model. For question-table encoder, we incorporate BERT to enhance the detection of reference and ellipsis, thus better resolve conversations. For action decoder, we adopt simple feed-forward neural networks to avoid heuristically designed features. We achieve competitive performance on the SequentialQA dataset. Further studies show the effectiveness of each part proposed in our approach.