Fiche de REBUFFEL Clement

Titre : Towards Controllable and Factual Data-to-Text Generation Models
Date : 5 juillet 2022
Heure: 14:00
Résumé : Data-to-Text Generation (DTG) concists in describing complex data using Natural Language such that humans can better understand it. Automated systems for DTG traditionally consist in a collection of complex rules, crafted using domain expertise. Recently, approaches based on Neural Networks (NN) have been proposed. NN alleviate the need for costly domain expertise and are able to produce fluent and in-domain descriptions. Unfortunately, descriptions written by NN tend to include incorrect statements, a behaviour called hallucination. This is mostly due to noisy data, where training descriptions do not completely align with their associated inputs, causing NN to learn spurious alignments and including fasle information in their outputs. This PhD tackles this issue along three axes. We've proposed 1) hierarchical models that better handle complex data; 2) training frameworks, so models can learn to from noisy examples; 3) an automated metric for semantic evaluation in DTG
Lieu : Jussieu, laboratoire ISIR ou salle SCAI