Query Generation for Answering Complex Questions in Russian Using a Syntax Parser
D. A. Evseev · Scientific and Technical Information Processing · 2022
Abstract This article presents a system that translates natural language questions into SPARQL queries. The question answering system includes a syntax parser that generates a parse tree of the input sentence; a component that generates a SPARQL query template based on the parse tree; and models that identify entities and relations to be inserted into the SPARQL query template. Entity extraction and relation ranking is performed using BERT. Training BERT for a Russian-language question answering system is faced with the problem of an insufficient volume of available training data. To counter this issue, we investigate the possibility of training multilingual BERT pretrained on the LC-QUAD2.0 dataset to perform the tasks of entity extraction and relation ranking on a small amount of Russian-language samples from the RuBQ dataset. The proposed question answering system, as tested on the RuBQ dataset, outperforms the accuracy of previous approaches.