I am a PhD candidate in C.S. from Oct. 2021 with an INRIA grant currrently working under the supervision of Pr.
J-C Régin (I3S).
My topic is : Automatic generation of text under constraints for the detection and monitoring of
visual pathologies.
I am part of Biovision of project-team at Inria and Constraints and Applications
team at I3S lab.
My PhD research tackled a critical real-world problem: the limited availability of diverse, high-quality sentences for standardized vision screening tests like MNREAD. I addressed this challenge by developing a novel method for generating highly constrained texts that adhere to strict requirements.
By formulating the task as a discrete optimization problem and leveraging Multivalued Decision Diagrams (MDDs), I achieved an innovative solution: exhaustive generation of valid sentences without the need for complex searching. To refine the results and ensure natural language, I incorporated a language model (GPT).
This method stands out by effectively handling intricate constraints, including sentence length, display rules, and even grammatical nuances in both English and French. Compared to existing tests like MNREAD's limited set, my approach produces hundreds of diverse, valid sentences. This significant advancement expands the range of patient assessments while simultaneously unlocking the potential for adapting the method to various languages, ultimately enhancing the versatility and impact of MNREAD.
Beyond the specific application of vision screening, my research showcases the power of MDDs as a potent tool for constrained text generation across diverse domains. This paves the way for future applications that can benefit from text generation under constraints.
