Abstract:
Resource Description Framework (RDF) is the standard for representing structured
knowledge on the Web. It is based on entities such as facts, events, and the relationships between them. RDF verbalizers are important to generate good quality textual descriptions from such RDF data. Despite the signi cant work done for the English language, no efforts have been directed towards low-resource languages like the Arabic language. This work promotes the development of RDF data-to-text (D2T) generation systems for the Arabic language by introducing a new Arabic dataset (AraWebNLG). A comparative study between multiple sequence-to-sequence models is also presented while studying the transfer of knowledge from pre-trained language models (AraBERT, AraGPT2 and mT5) to overcome data limitations. The analysis involves numerical metrics
(BLEU and Perplexity scores) as well as task-specific metrics related to the accuracy of the content selection and fluency of the generated text. The results highlight the importance of pre-training on a large corpus of Arabic data as the AraBERT initialized model is the best performing among the others. Text-to-text pre-training using mT5 is also able to achieve competitive results even with multilingual weights.
