Semantic Parsing for Conversational Question Answering over Knowledge Graphs

Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 2507–2522 May 2-6, 2023 ©2023 Association for Computational Linguistics Semantic Parsing for Conversational Question Answering over Knowledge Graphs Laura Perez-Beltrachini1, Parag Jain1, Emilio Monti2, Mirella Lapata1 1 School of Informatics, University of Edinburgh 2 Amazon Alexa {lperez,parag.jain,mlap}@inf.ed.ac.uk, monti@amazon.co.uk Abstract In this paper, we are interested in developing semantic parsers which understand natural lan- guage questions embedded in a conversation with a user and ground them to formal queries over definitions in a general purpose knowl- edge graph (KG) with very large vocabularies (covering thousands of concept names and re- lations, and millions of entities). To this end, we develop a dataset where user questions are annotated with SPARQL parses and system an- swers correspond to execution results thereof. We present two different semantic parsing ap- proaches and highlight the challenges of the task: dealing with large vocabularies, mod- elling conversation context, predicting queries with multiple entities, and generalising to new questions at test time. We hope our dataset will serve as useful testbed for the development of conversational semantic parsers.1 1 Introduction Conversational information seeking is the process of acquiring information through conversations (Za- mani et al., 2022). Recent years have seen an in- creasing number of applications aiming to build conversational interfaces based on information re- trieval (Radlinski and Craswell, 2017) and user recommendation (Jannach et al., 2021). The pop- ularity of intelligent voice assistants such as Ama- zon’s Alexa or Apple’s Siri has further stimulated research on question answering over general pur- pose knowledge graphs (e.g., Wikidata). Key to question answering in this context is the ability to ground natural language onto concepts, entities, and relations in order to produce an executable query (e.g., SPARQL) which will retrieve an answer or denotation from the knowledge graph (KG). This grounding process, known as semantic pars- ing has been studied in the context of one or few domain-specific databases (Yu et al., 2019a; Jain 1Our dataset and models are released at SPICE. and Lapata, 2021; Suhr et al., 2018) or without taking the conversational nature of the task into ac- count (Reddy et al., 2014; Yih et al., 2016; Dubey et al., 2019; Gu et al., 2021). However, due to the complexities of the semantic parsing task, there are no large scale datasets consisting of informa- tion seeking conversations with executable queries against a KG. Conversational semantic parsing over KGs requires handling very large vocabular- ies covering thousands of concept names and rela- tions, and millions of entities rather than special- ized terms consisting of hundreds of tables and column names. Moreover, information seeking conversations are by nature incremental involving interrelated rather than isolated questions. In this work, we create SPICE, a Semantic ParsIng dataset for Conversational quEstion an- swering over Wikidata. SPICE consists of user- assistant interactions where natural language ques- tions are paired with SPARQL parses and answers provided by the system correspond to SPARQL exe- cution results. We derive this dataset …

Content truncated to 500 words