A Corpus and Semantic Parser for Multilingual Natural Language Querying of OpenStreetMap

Proceedings of NAACL-HLT 2016, pages 740–750, San Diego, California, June 12-17, 2016. c⃝2016 Association for Computational Linguistics A Corpus and Semantic Parser for Multilingual Natural Language Querying of OpenStreetMap Carolin Haas Computational Linguistics Heidelberg University 69120 Heidelberg, Germany haas1@cl.uni-heidelberg.de Stefan Riezler Computational Linguistics & IWR Heidelberg University 69120 Heidelberg, Germany riezler@cl.uni-heidelberg.de Abstract We present a corpus of 2,380 natural language queries paired with machine readable formu- lae that can be executed against world wide geographic data of the OpenStreetMap (OSM) database. We use the corpus to learn an ac- curate semantic parser that builds the basis of a natural language interface to OSM. Fur- thermore, we use response-based learning on parser feedback to adapt a statistical machine translation system for multilingual database access to OSM. Our framework allows to map fuzzy natural language expressions such as “nearby”, “north of”, or “in walking distance” to spatial polygons on an interactive map. Fur- thermore, it combines syntactic complexity and compositionality with a reasonable lexical variability of queries, making it an interesting new publicly available dataset for research on semantic parsing. 1 Introduction OpenStreetMap1 (OSM) is a community-built database of geographic data, containing user- contributed local and up-to-date information about landmarks all over the world. Currently, the database contains over 3 billion data objects and is continuously growing with contributions from over 2 million registered users. While the main API is op- timized for editing map data, there exists an API that allow to filter map data based on search criteria such as location, type of objects, or features with which objects are tagged. However, issuing a query that is 1http://www.openstreetmap.org executable against the OSM database still requires detailed knowledge of database internals, something that cannot be expected from a layman user. The goal of our work is the development of an interface to OSM that lets a user ask a question in natural language, which is then parsed into a database query that is executable against a web- based filtering tool and returns OSM data on an interactive map. For example, we want a user without detailed knowledge of OSM to be able to ask questions that embrace the “fuzziness” of nat- ural language, for example, “What are the loca- tions, names and telephone numbers of hotels in Paris with wheelchair access that are close to the station Gare du Nord?”. To find such informa- tion one would have to issue a query that requires detailed knowledge of the database and the query language: “area[name=‘Paris’]→.a;node(area.a) [name=‘Gare du Nord’]→.b;node(around.b:1000) [tourism=‘hotel’][wheelchair=‘yes’];out;”. Addi- tionally, we present an adaptation of a statistical ma- chine translation (SMT) system for multilingual ac- cess to OSM by response-based learning from parser executability of translated queries. As a starting point for our natural language in- terface we built a corpus of 2,380 natural lan- guage queries paired with machine readable lan- guage (MRL) formulae that we used to extract a se- mantic parser. We chose to manually creating a cor- pus of MRLs from which structure and weights of a …

Content truncated to 500 words