This poster describes a new approach to constrained text generation using NLP and CP-AI-OR reasoning. The poster argues that beam search fails to generate constrained text when the solution space is highly constrained, and that a Constraint Programming-based approach can overcome this limitation. The poster presents a case study based on the MNREAD test, a psychophysical test based on standardized sentences. It shows that a Multi-valued Decision Diagram model can generate sentences that satisfy the MNREAD rules. The poster also discusses the advantages of their approach, including its modularity, flexibility, and ability to consider constraints at the generation stage. Finally, the authors discuss the perspectives of their work, including the possibility of bridging CP and ML.