Understanding Search Queries in Natural Language

NEVĚŘILOVÁ, Zuzana and Matej KVAŠŠAY. Understanding Search Queries in Natural Language. In Horák, Aleš and Rychlý, Pavel and Rambousek, Adam. Proceedings of Recent Advances in Slavonic Natural Language Processing, RASLAN 2018. Brno: Tribun EU, 2018, p. 85-93. ISBN 978-80-263-1517-9.

Basic information

• Original name: Understanding Search Queries in Natural Language
• Authors: NEVĚŘILOVÁ, Zuzana (203 Czech Republic, guarantor, belonging to the institution) and Matej KVAŠŠAY (703 Slovakia).
• Edition: Brno, Proceedings of Recent Advances in Slavonic Natural Language Processing, RASLAN 2018, p. 85-93, 9 pp. 2018.
• Publisher: Tribun EU

Other information

• Original language: English
• Type of outcome: Proceedings paper
• Field of Study: 10201 Computer sciences, information science, bioinformatics
• Country of publisher: Czech Republic
• Confidentiality degree: is not subject to a state or trade secret
• Publication form: printed version "print"
• WWW: URL
• RIV identification code: RIV/00216224:14330/18:00109726
• Organization unit: Faculty of Informatics
• ISBN: 978-80-263-1517-9
• ISSN: 2336-4289
• UT WoS: 000612420300011
• Keywords (in Czech): search intent; search query parsing
• Keywords in English: search intent; search query parsing

Abstract

This work is part of a project aiming to provide one single search endpoint for all company data. We present a search query parser that takes a speech-to-text output, i.e. a sentence. The output is a structured representation of the search query from which a SPARQL query is generated. The SPARQL is then applied to an ontology with the company data. The parsing procedure consists of two steps. First, the search intent is detected, second, the query is parsed based on the search intent. For the intent classification, we use word embeddings with boosting of top 5 words, and support vector machines. For the parsing, we use semantic role labeling, named entity recognition, and external resources such as ConceptNet and DBPedia. The final parsing step is rule-based and related to the ontology structure. The intent classifier accuracy is 94%. In the subsequent manual evaluation,the resulting structures were complete and correct in 51% cases, in 34.57% of cases it was complete and correct but it also contained irrelevant information.

Links

• EF16_013/0001781, research and development project: Name: LINDAT/CLARIN - Výzkumná infrastruktura pro jazykové technologie - rozšíření repozitáře a výpočetní kapacity