Jointly Learning Conversational Semantic Parsing and Answerability Detection

Anonymous

Keywords: Semantic Parsing, Answerability, Pretrained Language model, Conversational System

TL;DR: Jointly learn conversational semantic parsing and answerability detection using a pretrained sequence to sequence model

Abstract: Conversational semantic parsing is a challenging task that aims to automatically translate user utterances into logic forms (e.g., SQL queries) in multi-turn interactions. Most existing conversational semantic parsing models handle this task by assuming the user utterances are well-formed and answerable. Although these models have achieved prompting results on the Text2SQL task, few methods consider the answerability detection problem, causing the conversational semantic parser not able to deal with the practical scenario. To fill this gap, we propose to jointly learn the conversational semantic parsing and the answerability detection task on top of the pretrained sequence to sequence model. In this way, the model would be able to detect the answerability of the user utterance, respond with the translated SQL query for the answerable questions, and generate clarification answer for the unanswerable and ambiguous questions. Experimental results show that our joint learning framework performs satisfactorily for the answerability detection task, and results in performance improvements in terms of the generated SQL quality.