We focus on the task of querying a knowledge graph (KG) using natural language. KG can be queried reliably through formal languages such as SPARQL; however, this requires a complex translation from the natural language to the formal language. Large Language Models (LLMs) are able to directly answer natural language questions but provide no guarantees regarding the validity of the answers they generate. We propose a neurosymbolic system, called AlasQA, which answers natural language questions by combining the reliability of a formal language like SPARQL with the power of LLMs. The approach relies on an intermediate tool for interactive SPARQL query construction. Experiments conducted on the QALD and TEXT2SPARQL datasets validate the relevance of this hybrid method.