Authors:
Anicet Ondo
1
;
Laurence Capus
1
and
Mamadou Bousso
2
Affiliations:
1
Dep. of Computer and Software Engineering, Laval University, Quebec, G1V 0A6, Quebec, Canada
;
2
Dep. Computer Science, Iba Der Thiam University, Thies, 967, Senegal
Keyword(s):
Question-Answering System, SPARQL Language, Transformer, Neural Network, Natural Language Processing.
Abstract:
A well-designed ontology must be capable of addressing all the needs it is intended to satisfy. This complex task involves gathering all the potential questions from future users that the ontology should answer in order to respond precisely to these requests. However, variations in the questions asked by users for the same need complicate the interrogation process. Consequently, the use of a question-answering system seems to be a more efficient option for translating user queries into the formal SPARQL language. Current methods face significant challenges, including their reliance on predefined patterns, the quality of models and training data, ontology structure, resource complexity for approaches integrating various techniques, and their sensitivity to linguistic variations for the same user need. To overcome these limitations, we propose an optimal classification approach to classify user queries into corresponding SPARQL query classes. This method uses a neural network based on
Transformer encoder-decoder architectures, improving both the understanding and generation of SPARQL queries while better adapting to variations in user queries. We have developed a dataset on estate liquidation and Python programming, built from raw data collected from specialist forums and websites. Two transformer models, GPT-2 and T5, were evaluated, with the basic T5 model obtaining a satisfactory score of 97.22%.
(More)